| # | Dimension | pi (v1.0.2) | aider (v0.86.2) | Cline (v4.1.22) | Goose (v1.53.0) | Codex CLI (rust-v0.160.0) | OpenHands (v1.53.0) | OpenCode (v1.18.34) | Hermes (v2026.9.24) | OpenClaw (v2026.9.9) | Gemini CLI (v0.63.0) | SWE-agent (v1.1.0) | Qwen Code (v0.25.0) | DGM (a565fd2) | Autohand Code CLI (v0.9.8) | DeepSeek Harness (dsh-v0.2.1-alpha.3) | Deep Agents (v0.7.25) | Grok Build (2bdd1d6) | Harness² (1cfe254) | Agensh (v1.0) | dsh-superagent (v0.1.0 @ 6337fc8) | ZCode (3.16.2 @ 3c5b227) | Everruns (v0.49.0 @ 4e8002a) | MIKE (v0.1.0 @ 382ec61) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Agent loop | Event-sourced; inner (tool/steer) + outer (follow-up) loops; no iteration cap <sup>[5](pi/article.md#user-content-fn-5), [6](pi/article.md#user-content-fn-6), [10](pi/article.md#user-content-fn-10)</sup> | No tool-call loop; parse→apply→reflect over text edit formats; REPL + reflection (≤3) + retry; no REPL iteration cap <sup>[4](aider/article.md#user-content-fn-4), [6](aider/article.md#user-content-fn-6), [8](aider/article.md#user-content-fn-8)</sup> | Host-independent SDK `AgentRuntime.execute`; turn loop + tool execution; `maxIterations: undefined` in VS Code host — bound is 5 identical tool calls <sup>[7](cline/article.md#user-content-fn-7), [13](cline/article.md#user-content-fn-13), [14](cline/article.md#user-content-fn-14)</sup> | Two loops, migration in progress: legacy `agent.rs` plain `loop{}` (default, `max_turns`=1000) vs opt-in effect-sourced state machine (`GOOSE_STATE_MACHINE=1`); LLM call dead last <sup>[3](goose/article.md#user-content-fn-3), [6](goose/article.md#user-content-fn-6), [7](goose/article.md#user-content-fn-7), [10](goose/article.md#user-content-fn-10)</sup> | **Three nested**: `RegularTask::run` → `run_turn` `loop{}` (one sampling request/iter) → stream-consumer + retry; condition-driven; **no max_turns anywhere in 154 crates** — bound is retry budget + compaction-once-per-step + user interrupt <sup>[5](codex/article.md#user-content-fn-5), [7](codex/article.md#user-content-fn-7), [9](codex/article.md#user-content-fn-9)</sup> | Event-sourced; `LocalConversation._run` → `Agent.step` → pure `classify_response`; `max_iteration_per_run` 500; stop hooks can veto FINISHED <sup>[18](openhands/article.md#user-content-fn-18), [19](openhands/article.md#user-content-fn-19), [20](openhands/article.md#user-content-fn-20), [26](openhands/article.md#user-content-fn-26)</sup> | `runLoop` `while(true)` in `session/prompt.ts`; exit = finish reason ≠ `tool-calls` + no pending tool calls + `parentID` == user message; `maxSteps` defaults `Infinity` (injects wrap-up prompt, never kills); `prompt()` serializes concurrent prompts via `state.ensureRunning` <sup>[5](opencode/article.md#user-content-fn-5), [6](opencode/article.md#user-content-fn-6), [7](opencode/article.md#user-content-fn-7), [8](opencode/article.md#user-content-fn-8)</sup> | Synchronous one-turn `run_conversation` (turn loop lives in gateway/CLI drivers); in-turn loop, `max_iterations` unlimited by default — `sys.maxsize` (subagents 250 via `delegation.max_iterations`); per-phase sibling modules (preflight, assembly, tool round, overflow, compaction, stop gates, finalizer); session turn lease serializes concurrent prompts <sup>[5](hermes/article.md#user-content-fn-5), [6](hermes/article.md#user-content-fn-6), [7](hermes/article.md#user-content-fn-7), [28](hermes/article.md#user-content-fn-28)</sup> | Four nested layers: `runLoop()` raw model→tool loop (executes tool calls as they stream in via `streamAgentResponse`, steering-queue injection) → `Agent` class (transcript ownership) → `AgentSession` mixin chain → gateway execution; no hard max-turn counter — delegated to `prepareNextTurn`/`shouldStopAfterTurn` hooks; tool-loop recovery detector <sup>[7](openclaw/article.md#user-content-fn-7), [9](openclaw/article.md#user-content-fn-9), [10](openclaw/article.md#user-content-fn-10), [11](openclaw/article.md#user-content-fn-11), [12](openclaw/article.md#user-content-fn-12)</sup> | `sendMessageStream` → `processTurn` → `Turn.run`; `MAX_TURNS`=100; per-turn composite model router (fallback → ... → default); loop detector + next-speaker LLM check; loop split: core streams/decides, CLI schedules/feeds back <sup>[6](gemini-cli/article.md#user-content-fn-6), [8](gemini-cli/article.md#user-content-fn-8), [12](gemini-cli/article.md#user-content-fn-12), [13](gemini-cli/article.md#user-content-fn-13)</sup> | Sync `while not done` - one model call → one command per step, `.traj` rewritten every step; termination = submit/exit/forfeit/error-autosubmit/cost/context/timeout; **no step cap**; optional `RetryAgent` outer attempt loop <sup>[9](swe-agent/article.md#user-content-fn-9), [10](swe-agent/article.md#user-content-fn-10), [11](swe-agent/article.md#user-content-fn-11), [12](swe-agent/article.md#user-content-fn-12)</sup> | Recursive `sendMessageStream` re-entry (no `processTurn`); `MAX_TURNS`=100 + session-turn cap + session-token-limit gate; 8 `SendMessageType`s; next-speaker check skipped by default; `Stop`/`MessageDisplay` hook stages <sup>[8](qwen-code/article.md#user-content-fn-8), [10](qwen-code/article.md#user-content-fn-10), [13](qwen-code/article.md#user-content-fn-13), [18](qwen-code/article.md#user-content-fn-18)</sup> | Outer: fixed evolution driver — 80 gens × 2 children, ThreadPoolExecutor, 1.5h per-child timeout, compile filter, JSONL checkpoint. Inner: single-shot `forward()` → native tool loop, **no step cap**, bare-`except` termination <sup>[2](dgm/article.md#user-content-fn-2), [3](dgm/article.md#user-content-fn-3), [10](dgm/article.md#user-content-fn-10), [11](dgm/article.md#user-content-fn-11)</sup> | Bounded `for` loop (`runAgentReactLoop`); cap 100 iterations (1,000 executing a plan), exhaustion ending in a summary call rather than a kill; model-declared finish; loop/reflection repair guards ride along; steering queue (≤32 messages) drains only at iteration boundaries <sup>[10](autohand/article.md#user-content-fn-10), [11](autohand/article.md#user-content-fn-11), [13](autohand/article.md#user-content-fn-13), [16](autohand/article.md#user-content-fn-16)</sup> | The loop itself is a bundle plugin row (`ReactLoopAgent`); `kick()` runs turns that re-derive context from the append-only session log at each step; **no built-in turn budget** — a run ends when the model stops calling tools, on abort, or via an `agent/turn-stopping` hook veto; an advisory repeat-tool reminder fires at 3/5/8 identical calls <sup>[1](deepseek-harness/article.md#user-content-fn-1), [12](deepseek-harness/article.md#user-content-fn-12), [14](deepseek-harness/article.md#user-content-fn-14), [15](deepseek-harness/article.md#user-content-fn-15)</sup> | **No loop in this repo** — `create_deep_agent()` assembles a middleware stack into LangChain's `create_agent` graph, whose routing edge exits when the model stops calling tools; LangGraph Pregel executes it in super-steps, checkpointing after every step; the only budget is `recursion_limit: 9_999`, enforced as `GraphRecursionError` only when no stop condition is reached <sup>[1](deepagents/article.md#user-content-fn-1), [51](deepagents/article.md#user-content-fn-51), [53](deepagents/article.md#user-content-fn-53), [54](deepagents/article.md#user-content-fn-54)</sup> | Explicit `loop {}` in the shell session actor; one sampler round-trip per iteration; a turn ends on a response with no tool calls; **no default turn cap** (`max_turns` is an optional per-agent-definition field, default `None`); a stationarity guard nudges at 8 identical steps (4 for reads/plan) and hard-stops at 12 (8); two-pass compaction prefires at 85% of the window <sup>[4](grok-build/article.md#user-content-fn-4), [7](grok-build/article.md#user-content-fn-7), [8](grok-build/article.md#user-content-fn-8), [10](grok-build/article.md#user-content-fn-10)</sup> | No turn loop in-package — a budgeted search over harness trees: k shared base rollouts → 2k dimension-split proposals (INV procedure / DOM practitioner knowledge) → k compositions; parallel mode (independent draws and compositions) or sequential mode (checkpointed rounds c1..ck, all checkpoints are finals); the ledger asserts R = 4k rollouts / P = 3k sessions per task-mode <sup>[3](harness2/article.md#user-content-fn-3), [9](harness2/article.md#user-content-fn-9), [11](harness2/article.md#user-content-fn-11)</sup> | Two nested loops: Copilot CLI turns over ACP inside a prompt-carried cooperation loop (observe, CLAIM, build, verify, push/PR/self-merge, PATCH_SUMMARY, repeat); no scheduler or assignment in code; wall-clock termination bent by runner deadline posts and an idle watchdog <sup>[2](agensh/article.md#user-content-fn-2), [10](agensh/article.md#user-content-fn-10), [49](agensh/article.md#user-content-fn-49)</sup> | Borrowed turn loop (dsh Sessions, #15); the native loop is operational: poll -> Trigger -> Eval -> Ticket/Assignment, pumped by 2 s passes (`evaluateDue`) <sup>[18](dsh-superagent/article.md#user-content-fn-18), [19](dsh-superagent/article.md#user-content-fn-19), [20](dsh-superagent/article.md#user-content-fn-20)</sup> | Unbounded turn loop fed by one runtime command queue (prompts, steering, background/workflow results, goal continuations); per iteration: drain queue, micro/auto-compact (rapid-refill circuit breaker), MCP refresh, reminder assembly, one model step; a session Goal layer continues the loop across turns when idle <sup>[3](zcode/article.md#user-content-fn-3), [4](zcode/article.md#user-content-fn-4), [18](zcode/article.md#user-content-fn-18)</sup> | Reason–act cycle over a journaled log: a worker leases the session (durable actor model) and runs the whole turn with no queue hop between steps; 500-iteration default cap; sans-IO pure planner core <sup>[1](everruns/article.md#user-content-fn-1), [21](everruns/article.md#user-content-fn-21), [32](everruns/article.md#user-content-fn-32)</sup> | Per-task executor loop in the server: stream one completion, persist the turn, run its tool calls, repeat; no iteration cap — bounds are a consecutive-failed-turn limit (per-task, default 3), loop detection, an optional loop breaker <sup>[9](mike/article.md#user-content-fn-9), [13](mike/article.md#user-content-fn-13)</sup> |
| 2 | Tool system | 8 built-ins (4 active by default) + registry; TypeBox validation; sequential/parallel; nested calls <sup>[22](pi/article.md#user-content-fn-22), [27](pi/article.md#user-content-fn-27), [28](pi/article.md#user-content-fn-28)</sup> | None — model emits SEARCH/REPLACE blocks parsed by edit-format coders; shell suggestions run only with per-block confirmation <sup>[14](aider/article.md#user-content-fn-14), [18](aider/article.md#user-content-fn-18)</sup> | 9 built-ins (7 active in VS Code act mode) + MCP natives + team tools (SDK/CLI); sequential default, adjacent-parallel batching; 30s default timeout <sup>[12](cline/article.md#user-content-fn-12), [28](cline/article.md#user-content-fn-28), [29](cline/article.md#user-content-fn-29)</sup> | Entirely MCP-shaped: every tool is an MCP server (builtins = in-process tokio duplex); `ext__tool` namespacing; no non-MCP path <sup>[2](goose/article.md#user-content-fn-2), [21](goose/article.md#user-content-fn-21), [23](goose/article.md#user-content-fn-23)</sup> | Responses API function tools; `ToolRegistry` (`IndexMap`) + `ToolExecutor` trait; `ToolExposure` (`Deferred` + `tool_search`); concurrent dispatch (`FuturesOrdered`); **no built-in read_file/write_file/edit** — reads via `exec_command`, edits via `apply_patch` <sup>[10](codex/article.md#user-content-fn-10), [23](codex/article.md#user-content-fn-23), [24](codex/article.md#user-content-fn-24), [25](codex/article.md#user-content-fn-25)</sup> | Spec/registry split; SDK never imports `openhands-tools`; Pydantic validation; parallel batches in original order <sup>[8](openhands/article.md#user-content-fn-8), [47](openhands/article.md#user-content-fn-47), [54](openhands/article.md#user-content-fn-54)</sup> | `ToolRegistry.tools()` rebuilds per request for the exact model; `usePatch` swaps `apply_patch` for `edit`+`write` on GPT-family; `tool.definition` plugin hook rewrites definitions; MCP + plugin tools join dynamically; `websearch` gated to opencode provider <sup>[9](opencode/article.md#user-content-fn-9), [10](opencode/article.md#user-content-fn-10), [11](opencode/article.md#user-content-fn-11)</sup> | Import-time self-registration (schemas validated at registration); single `handle_function_call` funnel (bridge unwrap → middleware → plugin hooks → ACP approval → execute); Tool Search bridge collapses 100+ tools to `tool_search`/`tool_describe`/`tool_call`; MCP discovery deliberately deferred out of import (once blocked the gateway 120s) <sup>[14](hermes/article.md#user-content-fn-14), [15](hermes/article.md#user-content-fn-15), [16](hermes/article.md#user-content-fn-16), [17](hermes/article.md#user-content-fn-17)</sup> | `AnyAgentTool` registry + runtime factories; effective surface merges six sources (core, shell, channel, OpenClaw, plugin, Tool Search) then sandbox/profile/provider/sender/group/sub-agent policy filters; Tool Search materializes on demand — most tools never enter model context <sup>[29](openclaw/article.md#user-content-fn-29), [30](openclaw/article.md#user-content-fn-30)</sup> | 26 built-ins; `ToolRegistry.sortTools()` priority order (built-in → project → MCP); **no tool search** - every active tool serialized into `getFunctionDeclarations()` each turn; `wait_for_previous` injected; model-family declaration overrides <sup>[24](gemini-cli/article.md#user-content-fn-24), [25](gemini-cli/article.md#user-content-fn-25), [26](gemini-cli/article.md#user-content-fn-26), [27](gemini-cli/article.md#user-content-fn-27)</sup> | The ACI: YAML-declared commands → OpenAI function schemas / shell strings; container-side script implementations; 9 text parsers; blocklist-only permissions, zero approvals <sup>[3](swe-agent/article.md#user-content-fn-3), [6](swe-agent/article.md#user-content-fn-6), [8](swe-agent/article.md#user-content-fn-8), [20](swe-agent/article.md#user-content-fn-20)</sup> | ~54 built-in tool classes (+ goal/omni/workflow tools, ~60-70 potential); lazy mode-gated registration; `ToolSearchTool` + `shouldDefer` schema hiding - only a curated core set rides each request <sup>[31](qwen-code/article.md#user-content-fn-31), [32](qwen-code/article.md#user-content-fn-32), [33](qwen-code/article.md#user-content-fn-33), [34](qwen-code/article.md#user-content-fn-34)</sup> | Exactly two tools (bash, editor); `tool_info()` + `tool_function()` contract; bash's "persistent state"/"no internet" claims false in code; `editor` overwrites whole files; 10k truncation <sup>[18](dgm/article.md#user-content-fn-18)</sup> | 131 default tools in a hand-rolled schema subset, relevance-filtered per iteration by keyword/category heuristics; MCP servers namespaced into the registry with flattened schemas; agent-authored meta-tools via `create_meta_tool` behind a handler blocklist <sup>[18](autohand/article.md#user-content-fn-18), [19](autohand/article.md#user-content-fn-19), [21](autohand/article.md#user-content-fn-21), [22](autohand/article.md#user-content-fn-22)</sup> | Effect-registered, per-agent-scope registry; every call traverses `tools/pre-execute` waterfall listeners (veto or `ask`) and `tools/post-execute` listeners; no tool search — native mode sends every mounted schema; the PTC presentation collapses the wire surface to a single `run_code` tool whose sub-calls re-enter the full guarded pipeline; MCP is client-only <sup>[19](deepseek-harness/article.md#user-content-fn-19), [21](deepseek-harness/article.md#user-content-fn-21), [38](deepseek-harness/article.md#user-content-fn-38)</sup> | Middleware-provided: fs tools (`ls` / `read_file` / `write_file` / `edit_file` / `delete` / `glob` / `grep`) + `execute` (execution-capable backends only, filtered per request by backend capability) + the `task` delegation tool; caller tools pass through verbatim; no tool search — visibility via construction allowlists and profile `excluded_tools`; MCP arrives via external adapters as ordinary tools <sup>[21](deepagents/article.md#user-content-fn-21), [15](deepagents/article.md#user-content-fn-15), [16](deepagents/article.md#user-content-fn-16), [23](deepagents/article.md#user-content-fn-23)</sup> | Registry-declared tools in three toolsets (native, concise, hashline) plus in-tree Codex and OpenCode ports; per-agent allowlist/denylist filtering backed by a vendor-name compatibility mapping between tool naming schemes <sup>[16](grok-build/article.md#user-content-fn-16), [17](grok-build/article.md#user-content-fn-17)</sup> | None of its own — the substrate's tools, reached through a pinned OpenCode checkout run from source via bun or a versioned Codex dist; the package's own tool surface is the file-editing workspace granted to proposer sessions <sup>[60](harness2/article.md#user-content-fn-60), [43](harness2/article.md#user-content-fn-43)</sup> | Copilot's native tools + git CLI + curl, plus three stdio MCP servers (Gitea, Mattermost, board) wrapped by the notice shim; a 60-entry MCP deny list fences repo/file/channel administration; Copilot's own subagent tools are excluded — the team is the subagent mechanism <sup>[39](agensh/article.md#user-content-fn-39), [40](agensh/article.md#user-content-fn-40), [46](agensh/article.md#user-content-fn-46)</sup> | Router Session: Engine commands only under a read-only preset (no file writes, subagents or web); workers: six orchestration verbs plus dsh's full tool set <sup>[11](dsh-superagent/article.md#user-content-fn-11), [13](dsh-superagent/article.md#user-content-fn-13), [62](dsh-superagent/article.md#user-content-fn-62)</sup> | ~45 built-ins (Bash with a deep command-analysis family, file/web/todo/plan tools, subagents + coordinator messaging, a dynamic-workflow authoring family, Cron/OffPeak scheduling, a `js` repl, Skill, ToolSearch, GoalRead) + MCP servers initialized/refreshed in-loop; every tool entry carries schema, result budget, timeout, trace policy and a permission spec <sup>[10](zcode/article.md#user-content-fn-10), [6](zcode/article.md#user-content-fn-6)</sup> | Capability-provided tools (~60 documented capabilities); every tool declares MCP-style hints plus a `SideEffectClass` (`Pure`/`Idempotent`/`AtMostOnce`) driving approvals and crash reruns; MCP catalogs paged in full, searched, revealed within 32 tools / 64 KiB defaults <sup>[12](everruns/article.md#user-content-fn-12), [45](everruns/article.md#user-content-fn-45)</sup> | 13 core tools (files, bash, board, LSP) behind one dispatch table; plan mode swaps in a read-only bash; MCP per project opt-in; extension tools are data-declared HTTP calls <sup>[24](mike/article.md#user-content-fn-24), [46](mike/article.md#user-content-fn-46)</sup> |
| 3 | Model providers | ~42 behind one `StreamFn`; per-request key resolution <sup>[18](pi/article.md#user-content-fn-18), [29](pi/article.md#user-content-fn-29)</sup> | litellm behind a `Model` wrapper; 313-entry YAML + substring heuristics (model-aware, load-bearing); weak model for commits/summaries <sup>[25](aider/article.md#user-content-fn-25), [26](aider/article.md#user-content-fn-26)</sup> | 228 providers, one generic Vercel AI SDK adapter; exact-ID resolution; data-driven per-model behavior; ~200k-line generated catalog <sup>[6](cline/article.md#user-content-fn-6), [23](cline/article.md#user-content-fn-23)</sup> | Single `Provider` trait (`stream()` required); 36 static + 48 declarative JSON + custom; 8,160-entry models.dev snapshot (zstd-bundled, ETag hot-swap) + substring heuristics; no name validation <sup>[34](goose/article.md#user-content-fn-34), [35](goose/article.md#user-content-fn-35), [36](goose/article.md#user-content-fn-36)</sup> | **Responses-API-only** (`WireApi` single variant; `"chat"` rejected); 5 built-in providers (openai, 2× Bedrock, ollama, lmstudio); **8 auth modes**; `ModelProvider` trait; WS-first + HTTP/SSE fallback to `/responses`; retries 5 (cap 100) <sup>[3](codex/article.md#user-content-fn-3), [31](codex/article.md#user-content-fn-31)</sup> | LiteLLM; Chat Completions + Responses API; per-completion `RouterLLM`; agent self-swaps via `switch_llm` <sup>[40](openhands/article.md#user-content-fn-40), [55](openhands/article.md#user-content-fn-55), [84](openhands/article.md#user-content-fn-84), [86](openhands/article.md#user-content-fn-86)</sup> | Live ModelsDev catalog (fetched, cached with locking, refreshed) — never ships stale; dual runtime: Vercel AI SDK default (`streamText`) + opt-in native `@opencode-ai/llm` behind `OPENCODE_EXPERIMENTAL_NATIVE_LLM`; both converge on one `LLMEvent` stream <sup>[12](opencode/article.md#user-content-fn-12), [13](opencode/article.md#user-content-fn-13), [14](opencode/article.md#user-content-fn-14), [15](opencode/article.md#user-content-fn-15)</sup> | ~40 providers as lazily-discovered plugins; `ProviderProfile` with `chat_completions` / `codex_responses` API modes; reasoning rides `assistant_msg["reasoning"]` <sup>[24](hermes/article.md#user-content-fn-24)</sup> | `llm-core` transport-adapter contract (`stream`/`streamSimple` → normalized async-iterable); `packages/ai` registry maps API IDs to providers; per-vendor extensions declare auth, catalogs, thinking policies; `modelFallback` is last-resort chain resolution, never runtime failover <sup>[33](openclaw/article.md#user-content-fn-33), [34](openclaw/article.md#user-content-fn-34)</sup> | Gemini API; per-turn composite router (fallback → override → approval-mode → classifiers → default); routing auditable via `ModelRoutingEvent` telemetry; sequence-model stickiness across turns <sup>[8](gemini-cli/article.md#user-content-fn-8), [19](gemini-cli/article.md#user-content-fn-19), [42](gemini-cli/article.md#user-content-fn-42)</sup> | LiteLLM - any provider by model-name string; per-instance ($3.00) / total cost + call limits; API keys as `SecretStr` with `$ENV`/`:::` rotation; special `human`, `replay`, `instant_empty_submit` models <sup>[13](swe-agent/article.md#user-content-fn-13)</sup> | 6-auth-wire dispatch; 13-preset registry + setup wizard; **no per-turn router** - explicit `/model` selection resolved per call; models.dev bundled/live catalog; Qwen PKCE OAuth primary <sup>[49](qwen-code/article.md#user-content-fn-49), [50](qwen-code/article.md#user-content-fn-50), [55](qwen-code/article.md#user-content-fn-55), [52](qwen-code/article.md#user-content-fn-52)</sup> | Inner agent = Bedrock Claude 3.5 Sonnet (o3-mini for polyglot); diagnoser = o1-2024-12-17; ~20 models listed, narrow wiring; `MAX_OUTPUT_TOKENS = 4096`, temp 0.7 <sup>[12](dgm/article.md#user-content-fn-12)</sup> | 17 providers behind one `LLMProvider` interface; the factory defaults to OpenRouter with an unconfigured-provider fallback; a capability flag (`supportsNativeToolCalling`) switches between native tool calls and a text emission protocol — much of the loop's repair machinery exists to police the text path <sup>[44](autohand/article.md#user-content-fn-44), [20](autohand/article.md#user-content-fn-20)</sup> | All model traffic goes through one seam, `ctx.llm`, with provider adapters behind it — the official DeepSeek API or BYOK providers; provider-scoping is visible in code (the session-log upload field is added only by the official adapter, never for BYOK) <sup>[12](deepseek-harness/article.md#user-content-fn-12), [30](deepseek-harness/article.md#user-content-fn-30)</sup> | Any LangChain chat model via `init_chat_model` (`provider:model` strings); per-model harness profiles (Anthropic, OpenAI Codex, NVIDIA built in) and provider profiles shape the resolved kwargs; public registration APIs exist for both profile kinds <sup>[4](deepagents/article.md#user-content-fn-4), [41](deepagents/article.md#user-content-fn-41)</sup> | First-party two ways: session/OAuth traffic via `cli-chat-proxy.grok.com/v1`, API-key inference via `api.x.ai/v1`; three wire backends (chat completions, messages, responses) selected per catalog entry; BYOK custom models carry their own base URL, key, extra headers, and a required context-window field <sup>[12](grok-build/article.md#user-content-fn-12), [13](grok-build/article.md#user-content-fn-13), [11](grok-build/article.md#user-content-fn-11), [14](grok-build/article.md#user-content-fn-14)</sup> | Solver / improver / selector / judge as separate env-configured channels over two transports — an OpenCode channel (provider-prefixed model ids) or a Claude CLI channel (Vertex auth); the improver defaults to the solver model <sup>[31](harness2/article.md#user-content-fn-31), [36](harness2/article.md#user-content-fn-36)</sup> | The borrowed harness's provider: Copilot only in this release (any other AGENT_KIND is fatal), Copilot pinned at 1.0.70, gateway-configured model (defaults `gpt-5.6-sol`, high effort, 272k window) <sup>[14](agensh/article.md#user-content-fn-14), [6](agensh/article.md#user-content-fn-6)</sup> | None of its own: models are whatever the dsh host provides (#15); the bundle's only model-facing choices are its two presets, router and worker <sup>[2](dsh-superagent/article.md#user-content-fn-2)</sup> | A real registry: the bundled built-in config (revision 33) defines 21 provider templates over three API shapes (anthropic-messages, openai-chat-completions, openai-responses), GLM-first (Z.ai/BigModel, GLM-5.3) ahead of Kimi, MiniMax, DeepSeek, Qwen, MiMo, OpenAI, Anthropic, xAI, OpenRouter, OpenCode Go/Zen; built-in configs are revision-checked releases (bundled / last-known-good / remote-synced); official coding-plan gateway rewrites plan endpoints <sup>[27](zcode/article.md#user-content-fn-27), [28](zcode/article.md#user-content-fn-28), [30](zcode/article.md#user-content-fn-30)</sup> | Provider-neutral driver layer: OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI, Mistral, Fireworks, Cloudflare, Meta, MAI and others; switching provider does not change the agent definition; hosted tools/search negotiated per driver <sup>[14](everruns/article.md#user-content-fn-14)</sup> | Named OpenAI-compatible endpoints (llama.cpp home case) behind one streaming client; per-provider concurrency slots; agent extensions join as `agent://` providers; `llm` extensions reshape requests <sup>[35](mike/article.md#user-content-fn-35), [54](mike/article.md#user-content-fn-54)</sup> |
| 4 | Prompt construction | Structured sections, diffed per turn, XML-tagged <sup>[15](pi/article.md#user-content-fn-15), [16](pi/article.md#user-content-fn-16)</sup> | Fixed wire order with synthetic user/assistant pairs; prompt-cache breakpoints on stable prefixes; repo map re-injected per turn <sup>[10](aider/article.md#user-content-fn-10), [11](aider/article.md#user-content-fn-11)</sup> | Template + placeholder replacement; `MessageBuilder` normalization; rules as `# Rules` sections <sup>[21](cline/article.md#user-content-fn-21)</sup> | `PromptManager` from extension info + directory hints + goose mode; toolshim/native split per model; moim agent-only turn-context message <sup>[14](goose/article.md#user-content-fn-14)</sup> | `Prompt` → `ResponsesApiRequest` (`store: false`, `stream: true`, thread `prompt_cache_key`); **responses-lite** shape (UUIDv5 prefix-cache IDs); assembly base→developer→user; fragments hard-capped (AGENTS.md ≤10K) <sup>[14](codex/article.md#user-content-fn-14), [32](codex/article.md#user-content-fn-32)</sup> | Cached incremental view with enforced invariants; `<available_skills>` progressive disclosure <sup>[37](openhands/article.md#user-content-fn-37), [38](openhands/article.md#user-content-fn-38), [15](openhands/article.md#user-content-fn-15)</sup> | System prompt assembled per turn (environment, instructions, MCP servers, skills); `SessionProcessor.create()` pre-captures the working tree before the model call; `handle.process()` calls `LLM.Service` <sup>[25](opencode/article.md#user-content-fn-25), [26](opencode/article.md#user-content-fn-26), [27](opencode/article.md#user-content-fn-27)</sup> | Byte-stable system prompt for the conversation's life (prefix-cache invariant); skill invocations ride user messages, never the system prompt; tool schemas memoized per registry generation <sup>[8](hermes/article.md#user-content-fn-8)</sup> | Context-engine assembly per turn (history from per-agent SQLite, replay validated, context-window limited); skills delivered through the prompt system (`skillsSnapshot`, instruction-delivery caches); session-bound <sup>[26](openclaw/article.md#user-content-fn-26), [40](openclaw/article.md#user-content-fn-40)</sup> | System prompt + tiered GEMINI.md injection (global → extension/project → per-directory JIT); per-turn model-dependent tool refresh; `BeforeModel` hooks can rewrite the request <sup>[46](gemini-cli/article.md#user-content-fn-46), [19](gemini-cli/article.md#user-content-fn-19), [20](gemini-cli/article.md#user-content-fn-20)</sup> | Jinja2 system/instance/next-step templates + auto-generated `command_docs`; 100k observation cap with elision note; history processors (identity/trim/cache-control); no summarization - elide/trim/tag only <sup>[14](swe-agent/article.md#user-content-fn-14), [15](swe-agent/article.md#user-content-fn-15)</sup> | 6-layer system prompt (`autoMemory` last, cache-prefix aware); frozen at `startChat`, refreshed on events; Anthropic two-block `cache_control` split; `<system-reminder>` prelude re-prepended after compaction <sup>[57](qwen-code/article.md#user-content-fn-57), [58](qwen-code/article.md#user-content-fn-58), [59](qwen-code/article.md#user-content-fn-59), [60](qwen-code/article.md#user-content-fn-60)</sup> | Diagnoser prompt = full inner-agent source + failure evidence + **private test patch** → JSON contract; rewrite prompt = summary + suggestion + problem description; guardrails are prompt text <sup>[13](dgm/article.md#user-content-fn-13), [14](dgm/article.md#user-content-fn-14), [15](dgm/article.md#user-content-fn-15)</sup> | System prompt rebuilt per request by `SystemPromptBuilder`, including a `## User Preferences & Memory` block recalled from the event-sourced store plus an extractive summary tree; skills resolved across 11 sources with later-sources-win precedence <sup>[33](autohand/article.md#user-content-fn-33), [8](autohand/article.md#user-content-fn-8)</sup> | Context is derived per step from the session log under a stated invariant (model-visible ⟺ logged); skill summaries are advertised in the prompt with bodies fetched on invocation; threshold compaction emits a `<compacted-summary>` keeping the newest 16% of the window verbatim <sup>[5](deepseek-harness/article.md#user-content-fn-5), [37](deepseek-harness/article.md#user-content-fn-37), [17](deepseek-harness/article.md#user-content-fn-17)</sup> | Caller prompt + harness-profile suffix (the authored base prompt is deprecated); middleware appends fragments per request — the subagent roster inside the `task` tool description, skills metadata under progressive disclosure, memory files under `<agent_memory>`; LangChain prompt-caching middleware preserves the prefix <sup>[16](deepagents/article.md#user-content-fn-16), [42](deepagents/article.md#user-content-fn-42), [39](deepagents/article.md#user-content-fn-39), [7](deepagents/article.md#user-content-fn-7)</sup> | Per-turn assembly from the agent definition, project instructions, and filtered tool schemas; at the compaction threshold, summaries replace history — the threshold itself resolves through the remote-settings tiers to a hardcoded 85% default, and the remote payload can also carry a server-decided consent-gate notice <sup>[10](grok-build/article.md#user-content-fn-10), [3](grok-build/article.md#user-content-fn-3)</sup> | Evidence-as-files plus a short mission pointer; trajectories full and untruncated (per-observation clip 50k default, 10k floor enforced in code); per-dimension job templates interpolated with BenchPrompts values that may not mention grading <sup>[16](harness2/article.md#user-content-fn-16), [24](harness2/article.md#user-content-fn-24), [29](harness2/article.md#user-content-fn-29)</sup> | One rendered file (`agents.txt` + brief, to `/work/AGENTS.md`) carries identity, substrates, board discipline, event semantics and the work cycle; per turn the dispatcher prepends a board snapshot as a single-use ACP prompt prefix; urgent notices are appended to tool results mid-turn <sup>[13](agensh/article.md#user-content-fn-13), [16](agensh/article.md#user-content-fn-16), [19](agensh/article.md#user-content-fn-19)</sup> | A shipped prompt set is the behaviour layer: `orchestrator.md` (16 lines), the `admission.md` rubric, `dispatch.md`, worker/work prompts and `prompts/team/*` norms; candidate cards rendered by the app <sup>[12](dsh-superagent/article.md#user-content-fn-12), [36](dsh-superagent/article.md#user-content-fn-36), [60](dsh-superagent/article.md#user-content-fn-60)</sup> | Per-turn assembly in the loop: system sections + injected reminders (collaboration mode, todo, goal state as an authoritative block with the objective wrapped as untrusted data, prefetched memory recall, plan mode); continuation turns get a generated prompt with budget accounting and a completion-audit checklist; verifier turns get a dedicated no-tools, JSON-only prompt <sup>[9](zcode/article.md#user-content-fn-9), [14](zcode/article.md#user-content-fn-14), [13](zcode/article.md#user-content-fn-13)</sup> | Folded from the session log each reason step: system prompt + journaled history (messages are a projection of events) + schemas as disclosed by tool search; compaction checkpoints shrink the prompt, never the log; execution phases annotated <sup>[5](everruns/article.md#user-content-fn-5), [28](everruns/article.md#user-content-fn-28)</sup> | Rebuilt per run from SQLite rows: base template with tool list, skills at container paths, latest summary, project `AGENTS.md` in `<instructions>`, mode templates, date, working directory <sup>[22](mike/article.md#user-content-fn-22), [23](mike/article.md#user-content-fn-23)</sup> |
| 5 | Memory/session | JSONL sessions; compaction with branch summarization <sup>[30](pi/article.md#user-content-fn-30)</sup> | In-memory cur/done messages; background-thread summarization; markdown log in repo; git auto-commit per turn <sup>[16](aider/article.md#user-content-fn-16), [27](aider/article.md#user-content-fn-27)</sup> | SQLite + file-backend; versioned whole-file JSON envelopes; automatic compaction by default plus manual `/compact` (was fake pre-CLINE-2503) <sup>[37](cline/article.md#user-content-fn-37), [45](cline/article.md#user-content-fn-45)</sup> | SQLite `sessions.db` + JSON blobs; full `Recipe` persisted on session record; explicit cross-session recall (chatrecall + SQL FTS); `user_visible`/`agent_visible` flags <sup>[40](goose/article.md#user-content-fn-40)</sup> | **Rollout JSONL** `~/.codex/sessions/YYYY/MM/DD/` (14 `RolloutItem` variants) + SQLite index; fork 3 modes (`Copied`/`CopiedDeferred`/`Referenced`); **cross-session LLM memory extraction** (`$CODEX_HOME/memories`); resume = client-side rebuild <sup>[33](codex/article.md#user-content-fn-33), [34](codex/article.md#user-content-fn-34)</sup> | JSONL `EventLog` per event; pluggable condensers; server `base_state.json` resume; no cross-session recall <sup>[49](openhands/article.md#user-content-fn-49), [60](openhands/article.md#user-content-fn-60)</sup> | SQLite via Drizzle/Effect (WAL); no in-memory conversation state across turns — working set re-read each iteration; per-iteration `compaction.isOverflow` → summarize head, keep token-budgeted tail; `compaction.prune()` forked in background <sup>[20](opencode/article.md#user-content-fn-20), [21](opencode/article.md#user-content-fn-21)</sup> | SQLite/WAL `SessionDB` (~30 mixins: sessions, messages, FTS5, titles, timeline); FTS5 with custom CJK bigram tokenizer; `session_search` cross-session recall; memory as provider ABC (builtin, honcho, hindsight) <sup>[21](hermes/article.md#user-content-fn-21), [22](hermes/article.md#user-content-fn-22), [23](hermes/article.md#user-content-fn-23), [29](hermes/article.md#user-content-fn-29)</sup> | Two SQLite DBs via `node:sqlite` (no external server): shared state DB (append-only `session_state_events`, per-agent cursors) + per-agent DB (`session_nodes` → `session_windows` → `transcript_events`, append-only, compressed blobs, FTS); layered memory — `MEMORY.md` seed, vector/keyword index, active-memory proactive surfacing, dreaming/curation <sup>[35](openclaw/article.md#user-content-fn-35), [36](openclaw/article.md#user-content-fn-36)</sup> | JSONL session recording under `~/.gemini/tmp/<projectHash>/chats/`; legacy compression (0.5 threshold, `<state_snapshot>` two-pass probe) + opt-in `ContextManager`; 4-tier GEMINI.md memory; background session-mining extraction service <sup>[43](gemini-cli/article.md#user-content-fn-43), [44](gemini-cli/article.md#user-content-fn-44), [46](gemini-cli/article.md#user-content-fn-46), [48](gemini-cli/article.md#user-content-fn-48)</sup> | Per-step `.traj` JSON (trajectory steps, full history, info, replay config, env); no compaction, no cross-session memory; trajectories double as replay inputs and few-shot demos <sup>[16](swe-agent/article.md#user-content-fn-16), [22](swe-agent/article.md#user-content-fn-22)</sup> | JSONL `<uuid>.jsonl` under `~/.qwen/projects/<cwd>/chats`; 3-tier compaction (warn/auto 0.85/hard, in-place); QWEN.md + AGENTS.md + QWEN.local.md; `MemoryManager` (project/user/team layers) - per-query extraction + gated dream, no vector DB <sup>[62](qwen-code/article.md#user-content-fn-62), [63](qwen-code/article.md#user-content-fn-63), [66](qwen-code/article.md#user-content-fn-66), [68](qwen-code/article.md#user-content-fn-68)</sup> | The archive *is* the memory: run-id list, patch-only storage (`parent_commit` chain), per-node `metadata.json` + `self_evo.md` transcripts; per-generation JSONL checkpoint <sup>[5](dgm/article.md#user-content-fn-5), [6](dgm/article.md#user-content-fn-6), [3](dgm/article.md#user-content-fn-3)</sup> | Event-sourced JSONL memory (user + project scopes) with projection rebuild and 0.6-similarity dedup; an LLM extraction pass runs after every interactive turn; sessions are JSONL directories resumed by full replay, with a fresh-session fallback on replay failure <sup>[31](autohand/article.md#user-content-fn-31), [32](autohand/article.md#user-content-fn-32), [34](autohand/article.md#user-content-fn-34)</sup> | The append-only session event log (format v4) is the only state: per-session directories of checksummed zstd JSONL under `~/.dsh/sessions`, with torn-tail recovery; compaction triggers at ≈0.8 × window; **no built-in cross-session memory** — opt-in session search, or third-party memory over MCP <sup>[34](deepseek-harness/article.md#user-content-fn-34), [35](deepseek-harness/article.md#user-content-fn-35), [36](deepseek-harness/article.md#user-content-fn-36)</sup> | LangGraph checkpoints via a caller-supplied checkpointer (none by default); files live in a state channel via the default thread-scoped `StateBackend`, or Store/Filesystem/Composite backends; summarization is a non-mutating event with evicted history offloaded to `/conversation_history/*.md`; tool results are evicted at 20k tokens <sup>[40](deepagents/article.md#user-content-fn-40), [35](deepagents/article.md#user-content-fn-35), [34](deepagents/article.md#user-content-fn-34), [20](deepagents/article.md#user-content-fn-20)</sup> | Per-cwd session directories under `~/.grok/sessions`; the session-state archive is the chat history (`chat_history.jsonl`), enriched with git root and remote URL for traces; a separate memory store whose archive upload is skipped (with the skip recorded) when the session registry is disabled <sup>[41](grok-build/article.md#user-content-fn-41), [27](grok-build/article.md#user-content-fn-27), [37](grok-build/article.md#user-content-fn-37)</sup> | The improvement harness: a per-(run, mode, domain) library storing each adopted harness as its whole tree plus EXPERIENCE.md (why, per-edit effects, cue/lesson/evidence/confidence); a content-addressed session cache makes sweeps resumable <sup>[49](harness2/article.md#user-content-fn-49), [22](harness2/article.md#user-content-fn-22), [35](harness2/article.md#user-content-fn-35)</sup> | Team memory is the board: typed gists (100/300-char caps, detail behind a `G<n>` label for unfold), a 2,000-entry live view over fully retained MySQL history with boolean grep; claims expire (1h), patch summaries supersede, FAIL invalidates; individual memory is the persisted Copilot session plus archived HOME/workspace <sup>[8](agensh/article.md#user-content-fn-8), [32](agensh/article.md#user-content-fn-32), [34](agensh/article.md#user-content-fn-34)</sup> | One SQLite Store (WAL): Triggers, Threads, Decisions, Tickets, Assignments, receipts, cursors and a durable effect queue in the same file; Ticket stage derived, never stored <sup>[42](dsh-superagent/article.md#user-content-fn-42), [44](dsh-superagent/article.md#user-content-fn-44), [24](dsh-superagent/article.md#user-content-fn-24)</sup> | SQLite session store (messages, events, todos and the goal row) shared by all surfaces, with resume/fork; project memory is file-based (a `MEMORY.md` index under 200 lines over topic files) written by a sandboxed memory agent loop and consolidated by a dream pass, recalled by prefetch mid-turn; micro + auto compaction bounds context <sup>[17](zcode/article.md#user-content-fn-17), [25](zcode/article.md#user-content-fn-25), [26](zcode/article.md#user-content-fn-26)</sup> | The append-only PostgreSQL event log is the session: per-session sequence, messages as projections, resume as a fold of the tail; cross-session memory is file-shaped (organization Memories, agent/user scopes, Knowledge mounts); no wall-clock ceiling <sup>[6](everruns/article.md#user-content-fn-6), [28](everruns/article.md#user-content-fn-28)</sup> | The SQLite store is the session: every message, tool call and result a row, persisted before execution; threads fork with git checkpoints; rolling summarization at 90% of context; carry-over is files and the board <sup>[37](mike/article.md#user-content-fn-37), [39](mike/article.md#user-content-fn-39)</sup> |
| 6 | Reasoning/planning | Thinking levels forwarded; no planner/sub-agents (omitted) <sup>[29](pi/article.md#user-content-fn-29), [32](pi/article.md#user-content-fn-32)</sup> | Architect = sequential 2-model delegation with user gate; no autonomous sub-agents <sup>[24](aider/article.md#user-content-fn-24)</sup> | plan/act modes (differ by one tool); sub-agents + teams in SDK/CLI, disabled in VS Code <sup>[35](cline/article.md#user-content-fn-35), [38](cline/article.md#user-content-fn-38), [45](cline/article.md#user-content-fn-45)</sup> | GooseMode (Auto/Approve/SmartApprove/Chat); `summon` sub-agents (forced Auto, 25 turns, no nesting); recipes as declarative task programs; in-process cron schedules (always Auto) <sup>[29](goose/article.md#user-content-fn-29), [44](goose/article.md#user-content-fn-44), [46](goose/article.md#user-content-fn-46), [43](goose/article.md#user-content-fn-43)</sup> | **Sub-agents, two generations** (V1 `multi_agent_v1.*`, V2 plain names) via `multi_agent_version`, `agent_max_depth`; `update_plan` tool; **Guardian LLM auto-reviewer** in the approval chain <sup>[30](codex/article.md#user-content-fn-30), [22](codex/article.md#user-content-fn-22), [23](codex/article.md#user-content-fn-23)</sup> | Critic + iterative refinement gates termination; `/goal` LLM judge loop; sub-agent frontmatter defs <sup>[34](openhands/article.md#user-content-fn-34), [30](openhands/article.md#user-content-fn-30), [63](openhands/article.md#user-content-fn-63)</sup> | Subagents as full child sessions with nested `runLoop` (own history, own SQLite rows) via `task` tool; background tasks inject completion as synthetic user messages; no separate planner <sup>[22](opencode/article.md#user-content-fn-22), [23](opencode/article.md#user-content-fn-23)</sup> | No separate planner; subagents via `delegate_task` (250 iterations, tool grant carved from parent); post-turn background review fork proposes skill/memory writes; curator auto-transitions skill lifecycle on inactivity (7 days; pin/archive/consolidate, never delete) <sup>[9](hermes/article.md#user-content-fn-9), [11](hermes/article.md#user-content-fn-11), [13](hermes/article.md#user-content-fn-13), [26](hermes/article.md#user-content-fn-26)</sup> | Subagents via the `AgentHarness` pluggable interface (`runAttempt`) — gateway can dispatch turns to non-OpenClaw harnesses; tool-loop recovery detector; retry budgets (`BASE_RUN_RETRY_ITERATIONS` = 24); post-run `continue`/`handoff` directives <sup>[2](openclaw/article.md#user-content-fn-2), [12](openclaw/article.md#user-content-fn-12), [26](openclaw/article.md#user-content-fn-26)</sup> | No separate planner; subagents via `invoke_agent` (isolated registries, mandatory `complete_task`, no nesting); `maxTurns` 30 default; `write_todos` tracking <sup>[53](gemini-cli/article.md#user-content-fn-53), [56](gemini-cli/article.md#user-content-fn-56), [57](gemini-cli/article.md#user-content-fn-57), [24](gemini-cli/article.md#user-content-fn-24)</sup> | No planner, no sub-agents; model's own discussion text parsed as `thought`; `AskColleagues` samples N completions and debates; reviewer/chooser score whole *attempts* and retry <sup>[19](swe-agent/article.md#user-content-fn-19), [12](swe-agent/article.md#user-content-fn-12)</sup> | `agent` tool (legacy `task` alias); first-class Plan Mode (enter/exit, headless/ACP entry); Agent Teams (mailbox, task board, persisted teams, leader permission bridge); checkpointed resumable workflows; native LSP <sup>[75](qwen-code/article.md#user-content-fn-75), [76](qwen-code/article.md#user-content-fn-76), [77](qwen-code/article.md#user-content-fn-77), [78](qwen-code/article.md#user-content-fn-78)</sup> | No planner/sub-agents; fixed o1 diagnoser (never improved); the rewrite is done by the current parent agent — the one recursive mechanism; post-hoc −2..2 score gates nothing <sup>[13](dgm/article.md#user-content-fn-13), [20](dgm/article.md#user-content-fn-20)</sup> | Plan mode with code-enforced read-only gates and an accept modal; persistent goals (`goals.local.json`) with a continuation loop and self-attested, floor-checked completion (`provenance: 'reported'`); sub-agent loop copies (cap 10, depth 3); OS-process peer teams <sup>[47](autohand/article.md#user-content-fn-47), [29](autohand/article.md#user-content-fn-29), [37](autohand/article.md#user-content-fn-37), [17](autohand/article.md#user-content-fn-17)</sup> | No separate planner: the planning surface is tools from the generated 2,773-line catalog — `todo_write`, goal tools, `exit_plan_mode`, scheduling — plus a `subagent` tool with interrupt/list/send-message controls <sup>[20](deepseek-harness/article.md#user-content-fn-20)</sup> | No planner and no todo tool in the default stack — `write_todos` exists only in the OpenAI Codex profile (LangChain `TodoListMiddleware`); decomposition is delegated to subagents via `task`; optional `RubricMiddleware` grades the transcript with a grader sub-agent until satisfied or capped <sup>[22](deepagents/article.md#user-content-fn-22), [44](deepagents/article.md#user-content-fn-44), [16](deepagents/article.md#user-content-fn-16)</sup> | A todo tool and plan mode (Plan is one of the six permission modes) structure the work; decomposition leans on subagent delegation rather than a separate planner <sup>[18](grok-build/article.md#user-content-fn-18)</sup> | The search is the planning: contrast-driven editing (every edit traces to a trajectory moment; divergences first, then repeated failures, then risks), dimension-split proposals, composition over candidate pools, final diversity measured by pairwise line-Jaccard <sup>[8](harness2/article.md#user-content-fn-8), [7](harness2/article.md#user-content-fn-7), [10](harness2/article.md#user-content-fn-10)</sup> | No planner: emergent self-selection — the task issue is unassigned, workers open their own issues and reserve areas with expiring CLAIM gists, collisions settled by DM; verification is prompt-prescribed comparison against the reference, and board 'verified' status is a wording check <sup>[3](agensh/article.md#user-content-fn-3), [9](agensh/article.md#user-content-fn-9), [35](agensh/article.md#user-content-fn-35)</sup> | No planner: one Ticket per accepted request; the assessor judges admission only, decomposition inside the worker Session is dsh's (subagents under the standard preset) <sup>[20](dsh-superagent/article.md#user-content-fn-20), [28](dsh-superagent/article.md#user-content-fn-28)</sup> | No separate planner: TodoRead/Write and plan mode structure work; decomposition via subagents and agent-authored dynamic workflows; long-horizon planning is the Goal system — a persisted objective continued turn-over-turn, with completion judged by a separate verifier (evidence per requirement, no proxy signals), not by the working model <sup>[13](zcode/article.md#user-content-fn-13), [14](zcode/article.md#user-content-fn-14), [18](zcode/article.md#user-content-fn-18)</sup> | No separate planner: the reason–act loop plus session tasks/sub-agents, schedules and journaled timers; long-horizon work is long-lived sessions woken by timers, triggers and channels <sup>[19](everruns/article.md#user-content-fn-19)</sup> | Modes act / plan / explore (plan mounts the worktree read-only, forbids writing, requires approval before task creation); subtask trees — children before parents — are the decomposition; no separate planner <sup>[23](mike/article.md#user-content-fn-23)</sup> |
| 7 | Extensibility | TS extensions, hooks at every seam, MCP, skills <sup>[31](pi/article.md#user-content-fn-31)</sup> | 42 closed slash commands; model-settings YAML; no plugin API, MCP, or skills — configuration, not code <sup>[32](aider/article.md#user-content-fn-32), [33](aider/article.md#user-content-fn-33), [34](aider/article.md#user-content-fn-34)</sup> | Richest so far: `AgentRuntimePlugin`, 7-callback hooks, file hooks, skills, MCP, sub-agents <sup>[4](cline/article.md#user-content-fn-4), [36](cline/article.md#user-content-fn-36)</sup> | Broadest so far: MCP extension configs (4 transports), 12-event plugin hooks, recipes, schedules, sub-agents, skills, custom providers, the `Provider` trait <sup>[21](goose/article.md#user-content-fn-21), [42](goose/article.md#user-content-fn-42), [46](goose/article.md#user-content-fn-46), [43](goose/article.md#user-content-fn-43), [35](goose/article.md#user-content-fn-35), [34](goose/article.md#user-content-fn-34)</sup> | `ToolExecutor` trait + `ToolContributor`; **MCP wrapped** (`McpHandler`, client-only); skills = prompt injection; hooks (12 events, precedence #1 in approvals); plugins (manifest); **closed TUI slash commands**; user-declared providers <sup>[23](codex/article.md#user-content-fn-23), [26](codex/article.md#user-content-fn-26), [27](codex/article.md#user-content-fn-27), [28](codex/article.md#user-content-fn-28), [29](codex/article.md#user-content-fn-29), [31](codex/article.md#user-content-fn-31)</sup> | Skills, 3-kind hooks, plugins, MCP, sub-agents, marketplace, extensions, condenser, router, critic — 12 axes <sup>[70](openhands/article.md#user-content-fn-70), [72](openhands/article.md#user-content-fn-72), [78](openhands/article.md#user-content-fn-78)</sup> | `tool.definition` plugin hook; MCP servers contribute tools dynamically; skills assembled into the system prompt; subagents; permission rulesets; Effect `Tag` DI throughout <sup>[10](opencode/article.md#user-content-fn-10), [27](opencode/article.md#user-content-fn-27), [4](opencode/article.md#user-content-fn-4)</sup> | Providers as lazily-discovered plugins; `pre_tool_call` + fail-open `transform_tool_result` hooks in the dispatch funnel; memory provider ABC; runtime skill writes via `skill_manage` (`SKILL.md` + frontmatter, agentskills.io-compatible) <sup>[13](hermes/article.md#user-content-fn-13), [15](hermes/article.md#user-content-fn-15), [23](hermes/article.md#user-content-fn-23)</sup> | Capability-typed plugin system (`api.registerX` — providers, channels, embeddings, speech, media, web fetch, migrations); ~100 bundled extensions with `openclaw.plugin.json` manifests; lazy loading (Telegram `activation.onStartup: false`); skill authoring as first-class flow <sup>[3](openclaw/article.md#user-content-fn-3), [38](openclaw/article.md#user-content-fn-38), [39](openclaw/article.md#user-content-fn-39)</sup> | MCP servers (namespaced `mcp_` tools); skills (two-stage `activate_skill`); 11 hook types (command + runtime, can block/rewrite/replace chunks); extensions contribute servers/skills/agents/hooks/policies <sup>[60](gemini-cli/article.md#user-content-fn-60), [64](gemini-cli/article.md#user-content-fn-64), [49](gemini-cli/article.md#user-content-fn-49)</sup> | Config-driven: new ACI = bundle dir + YAML line; parser/processor/problem-statement protocols; 13-point `AbstractAgentHook` + `EnvHook`/`RunHook`; Pydantic `extra="forbid"` <sup>[18](swe-agent/article.md#user-content-fn-18)</sup> | MCP (least-changed); 22 hook types in 4 kinds (command/http/function/prompt); skills can register session-scoped hooks; Auto-Skills weekly curator; Claude Code plugin converter <sup>[71](qwen-code/article.md#user-content-fn-71), [72](qwen-code/article.md#user-content-fn-72), [73](qwen-code/article.md#user-content-fn-73), [80](qwen-code/article.md#user-content-fn-80)</sup> | New tool = file in `tools/`; new sampling method = one CLI string (bugged); policies `keep_all`/`keep_better`; subsets are JSON; prompt modules <sup>[18](dgm/article.md#user-content-fn-18), [4](dgm/article.md#user-content-fn-4), [8](dgm/article.md#user-content-fn-8), [16](dgm/article.md#user-content-fn-16)</sup> | Cross-vendor `SKILL.md` skills, including LLM-generated ones written to `.autohand/skills` with provenance metadata and hash-based refresh; a community registry guarded by a regex security scanner (install path only); extension packages run in-process through a trust-gated dynamic-import runtime <sup>[3](autohand/article.md#user-content-fn-3), [39](autohand/article.md#user-content-fn-39), [40](autohand/article.md#user-content-fn-40)</sup> | Cordis plugins for everything above the kernel — the loop included; an `apply` contract with schema-validated config, inject/provide wiring and effect disposers; layered YAML patch composition (bundles, profiles, presets, overlays); npm-installed profile plugins; Creator mode defines plugins at runtime inside a `node:vm` realm <sup>[6](deepseek-harness/article.md#user-content-fn-6), [7](deepseek-harness/article.md#user-content-fn-7), [39](deepseek-harness/article.md#user-content-fn-39), [41](deepseek-harness/article.md#user-content-fn-41)</sup> | LangChain middleware API: replace-by-name stack surgery with protected scaffolding (`FilesystemMiddleware` and `SubAgentMiddleware` cannot be excluded); custom backends via `BackendProtocol`; declarative, compiled, or remote (LangGraph Platform) subagents; Agent Skills with progressive disclosure; public profile registration <sup>[8](deepagents/article.md#user-content-fn-8), [9](deepagents/article.md#user-content-fn-9), [43](deepagents/article.md#user-content-fn-43), [41](deepagents/article.md#user-content-fn-41)</sup> | Custom agent definitions plus a feature registry whose resolution order recurs everywhere — requirements pin → env → config → remote settings → default; the remote tier is fetched from the proxy's `GET /v1/settings` and can even carry a remote sandbox kill switch <sup>[3](grok-build/article.md#user-content-fn-3), [10](grok-build/article.md#user-content-fn-10)</sup> | BenchmarkAdapter protocol + BenchSpec values (install mode, leak globs, subagent policy, deliverable serving); a new substrate needs a SurfaceSpec, install modes and a trajectory renderer; two prompt packs (adapt, tts) <sup>[55](harness2/article.md#user-content-fn-55), [58](harness2/article.md#user-content-fn-58)</sup> | Team size as a Compose-generator parameter (one worker/router pair per agent); tasks as ProgramBench IDs; the prompt set is file configuration treated as part of the measured apparatus; a new single-agent harness needs an ACP entrypoint/bridge pair — only Copilot ships <sup>[5](agensh/article.md#user-content-fn-5), [51](agensh/article.md#user-content-fn-51), [52](agensh/article.md#user-content-fn-52)</sup> | The bundle is the extension: a Cordis patch into dsh (3 plugin inserts, 2 presets, injected web client); new Sources via the `SourceKind` contract + docs/008 cookbook <sup>[2](dsh-superagent/article.md#user-content-fn-2), [5](dsh-superagent/article.md#user-content-fn-5), [53](dsh-superagent/article.md#user-content-fn-53)</sup> | A plugin marketplace: Official Marketplace (builtin-seeded + sha256-checked CDN zips), personal git/GitHub/URL/local sources, Claude-marketplace compatibility; manifests carry commands/agents/skills/hooks/MCP servers/config; UI plugins render interactive MCP Apps panels co-edited by user and agent (desktop-local at this pin); skills via bundles/plugins/lockfile; 7 lifecycle hook events behind a workspace trust/review flow <sup>[36](zcode/article.md#user-content-fn-36), [37](zcode/article.md#user-content-fn-37), [23](zcode/article.md#user-content-fn-23)</sup> | Capabilities are the extension unit (~60 built in, config-selected); MCP servers with live authorization; agent packages and GitHub-synced sources (two-way in 0.49.0); channels (Agent API, Slack, webhook, AG-UI, A2A); embeddable framework crate <sup>[3](everruns/article.md#user-content-fn-3), [50](everruns/article.md#user-content-fn-50)</sup> | Seven extension types (lsp, ui, css, i18n, agent, llm, tool), git-pinned; isolation by kind — opaque-origin iframes, per-extension gVisor sandboxes, wazero adapters, pure data; MCP pinned to one protocol era <sup>[41](mike/article.md#user-content-fn-41), [55](mike/article.md#user-content-fn-55)</sup> |
| 8 | Interfaces | TUI / print / RPC / SDK — all event subscribers <sup>[9](pi/article.md#user-content-fn-9)</sup> | prompt_toolkit CLI; `--message` one-shot; streamlit GUI; imperative rendering, no event stream <sup>[20](aider/article.md#user-content-fn-20)</sup> | VS Code webview over proto-bus gRPC (22 svcs/224 RPCs) + CLI host + npm SDK re-export <sup>[26](cline/article.md#user-content-fn-26), [39](cline/article.md#user-content-fn-39)</sup> | Rust CLI REPL + Electron desktop + ACP bridge + Telegram gateway; loop emits `AgentEvent`s; uniffi Python/Kotlin bindings via `goose-sdk` <sup>[20](goose/article.md#user-content-fn-20), [45](goose/article.md#user-content-fn-45), [41](goose/article.md#user-content-fn-41)</sup> | **All RPC**: TUI + `exec` + daemon are **app-server (JSON-RPC v2) clients**; `exec` is in-process; separate `exec-server` for remote/foreign-OS <sup>[2](codex/article.md#user-content-fn-2)</sup> | SDK / agent-server REST+WS / ACP host mode / TS client; all consume the event stream <sup>[6](openhands/article.md#user-content-fn-6), [66](openhands/article.md#user-content-fn-66)</sup> | SolidJS TUI / `serve` / ACP — all subscribe to the session bus (`Permission.Asked`, message/part updates); SQLite as the shared substrate <sup>[24](opencode/article.md#user-content-fn-24)</sup> | CLI + multi-tenant messaging gateway (Telegram/Discord/WhatsApp/Slack/Signal) with session-key namespaces and pairing-code flow + TypeScript TUI + ACP <sup>[4](hermes/article.md#user-content-fn-4), [20](hermes/article.md#user-content-fn-20)</sup> | **Channel-first**: ~25 channel adapters (Telegram built with `createChatChannelPlugin()`; WhatsApp, Discord, Signal, Slack) running in-process in the gateway; CLI and web UI are secondary surfaces <sup>[1](openclaw/article.md#user-content-fn-1), [20](openclaw/article.md#user-content-fn-20)</sup> | Ink TUI (interactive) + `gemini -p` non-interactive; ACP sessions; A2A server; programmatic SDK; VS Code IDE companion <sup>[67](gemini-cli/article.md#user-content-fn-67)</sup> | CLI subcommands (`run`, `run-batch`, `run-replay`, `traj-to-demo`, ...); textual `inspect` + web `inspector` trajectory viewers; Flask+SocketIO API + React frontend <sup>[17](swe-agent/article.md#user-content-fn-17)</sup> | Ink TUI; `qwen serve` ACP daemon (HTTP+SSE+WS) + `@qwen-code/acp-bridge`; 11 channel adapters spawning headless `qwen --acp`; web shell; Tauri desktop; VS Code companion; TS/Python/Java SDKs <sup>[83](qwen-code/article.md#user-content-fn-83), [84](qwen-code/article.md#user-content-fn-84), [85](qwen-code/article.md#user-content-fn-85), [86](qwen-code/article.md#user-content-fn-86), [88](qwen-code/article.md#user-content-fn-88)</sup> | CLI only; `analysis/visualize_archive.py` rebuilds the tree, regenerates eval commands <sup>[2](dgm/article.md#user-content-fn-2), [6](dgm/article.md#user-content-fn-6), [16](dgm/article.md#user-content-fn-16)</sup> | Ink TUI; one-shot/pipe CLI with JSON-schema output enforcement; ACP over ndJSON via an in-process `AgentSideConnection`; JSON-RPC over stdio with a defined method table for VS Code and external drivers <sup>[41](autohand/article.md#user-content-fn-41), [42](autohand/article.md#user-content-fn-42), [43](autohand/article.md#user-content-fn-43)</sup> | No interactive CLI TUI — a profile-boot CLI (`web` / `headless` / `sdk` / `sdk-minimal` / `acp`); the Web UI is the primary surface (default port 3080); headless one-shot prints the final answer on stdout; ACP is an automation-only subset; a JSON-RPC SDK server backs TypeScript and Python SDKs <sup>[42](deepseek-harness/article.md#user-content-fn-42), [43](deepseek-harness/article.md#user-content-fn-43), [44](deepseek-harness/article.md#user-content-fn-44), [45](deepseek-harness/article.md#user-content-fn-45)</sup> | Library first — the product is a compiled LangGraph graph, and invoke/stream are the runtime's; sibling packages: Deep Agents Code (`dcode`, client/server terminal app, TUI + headless), a `deepagents-acp` server, and Talon, a long-running host with messaging channels, cron, and SQLite checkpoints <sup>[45](deepagents/article.md#user-content-fn-45), [46](deepagents/article.md#user-content-fn-46), [47](deepagents/article.md#user-content-fn-47), [12](deepagents/article.md#user-content-fn-12)</sup> | Full-screen TUI; headless `-p` with JSON/streaming output; ACP over stdio plus a token-guarded local server; leader mode backed by relay and gateway WebSocket endpoints for remote control <sup>[12](grok-build/article.md#user-content-fn-12)</sup> | CLI suite only: harness2 (sweep), harness2-judge, -report, -setup, -check, -demo, -bridge; results as runs/ trees with summary/selection/judged JSON; a deterministic no-model demo exercises the full workflow <sup>[59](harness2/article.md#user-content-fn-59)</sup> | No team UI: Gitea, Mattermost, the board's HTTP/MCP surface, and the runner's CLI/Make lifecycle; operators watch the same Gitea/Mattermost the agents use; results are exported committed `main` HEAD plus sealed per-worker shutdown archives <sup>[37](agensh/article.md#user-content-fn-37), [44](agensh/article.md#user-content-fn-44)</sup> | Proactive ingress: GitHub (issues/PRs/reviews/CI/releases) and Feishu chats/DMs (disabled by default), both polled; egress is four gated reply kinds via the Outbox; a panel whose stages put 'needs a person' first <sup>[55](dsh-superagent/article.md#user-content-fn-55), [56](dsh-superagent/article.md#user-content-fn-56), [32](dsh-superagent/article.md#user-content-fn-32)</sup> | Three surfaces over one runtime: Electron desktop (updater, embedded browser, computer use), browser UI served by a local host (also `zcode --web`; token auth for non-local binds), terminal TUI/CLI under one `zcode` command; remote triggering from WeChat/Feishu/Telegram/webhook bots drives the same task flow; conversation sharing with a public-projection export <sup>[1](zcode/article.md#user-content-fn-1), [2](zcode/article.md#user-content-fn-2), [32](zcode/article.md#user-content-fn-32)</sup> | API-first and headless: management REST API + the stable Agent API (one agent per channel, agent keys, SSE, question/approval answers, per-caller spend limits); same shapes on framework `serve` hosting; REST for writes, SSE for reads <sup>[50](everruns/article.md#user-content-fn-50), [55](everruns/article.md#user-content-fn-55)</sup> | One surface: the Svelte 5 PWA (boards, chat, diff review, terminal) over REST + WebSocket from the same binary; a narrower bearer-token API; the PWA is also an offline client (IndexedDB mirror + write queue) <sup>[50](mike/article.md#user-content-fn-50), [51](mike/article.md#user-content-fn-51)</sup> |
| 9 | Failure handling | Auto-retry, truncation guards, abort; no iteration cap <sup>[21](pi/article.md#user-content-fn-21), [26](pi/article.md#user-content-fn-26), [10](pi/article.md#user-content-fn-10)</sup> | Exp-backoff retry (60s per-attempt cap); no retry on context exhaustion; malformed edits reflected (≤3); lint/test confirm loops <sup>[7](aider/article.md#user-content-fn-7), [18](aider/article.md#user-content-fn-18)</sup> | Provider retry 3×; output-limit recovery 3×; loop detection 3×/5× (reactive); mistake tracker 6; no host iteration cap <sup>[9](cline/article.md#user-content-fn-9), [14](cline/article.md#user-content-fn-14), [15](cline/article.md#user-content-fn-15)</sup> | Three-layer retry (provider pre-first-item ≤3, recipe `RetryManager`, micro-retries); `max_turns`=1000; repetition guard inert (registered `None`) — the bound is turns, not loops <sup>[7](goose/article.md#user-content-fn-7), [9](goose/article.md#user-content-fn-9)</sup> | Stream retries 5 (cap 100); WS→HTTP fallback resets counter; **unbounded connection retries** (5s→60s); **no loop detection by design**; tool errors become failure results, never break the loop <sup>[3](codex/article.md#user-content-fn-3), [9](codex/article.md#user-content-fn-9), [10](codex/article.md#user-content-fn-10)</sup> | Tenacity retries; condensation recovery; stuck detector; corrective nudges; hook fail-open <sup>[56](openhands/article.md#user-content-fn-56), [60](openhands/article.md#user-content-fn-60), [61](openhands/article.md#user-content-fn-61), [73](openhands/article.md#user-content-fn-73)</sup> | LSP diagnostics appended to `edit`/`write` output (errors only, cap 20/file) — model self-corrects next turn, no harness retry logic; permission deny fails fast; shutdown rejects pending asks closed <sup>[16](opencode/article.md#user-content-fn-16), [17](opencode/article.md#user-content-fn-17), [18](opencode/article.md#user-content-fn-18), [19](opencode/article.md#user-content-fn-19)</sup> | Overflow vs truncation as separate recovery verdicts; turn-start proactive compaction; identical-call loop guardrails; denial breaker counts approval refusals; curator never deletes, only archives <sup>[25](hermes/article.md#user-content-fn-25), [18](hermes/article.md#user-content-fn-18), [12](hermes/article.md#user-content-fn-12)</sup> | SQLite ingress queue with claim leases, per-lane serialization, retry, dead-letter, stall watchdog (at-least-once ordered delivery); auto-retry with exponential backoff (rate-limit/server errors — context overflow routes to compaction instead); reply dead-letter + reconciliation <sup>[24](openclaw/article.md#user-content-fn-24), [17](openclaw/article.md#user-content-fn-17), [27](openclaw/article.md#user-content-fn-27)</sup> | Typed `Retry`/`Error` turn events; loop detector aborts repeated patterns; hook failures non-fatal (fail-open); taint-risk escalation forces ASK_USER on build-file edits <sup>[9](gemini-cli/article.md#user-content-fn-9), [18](gemini-cli/article.md#user-content-fn-18), [51](gemini-cli/article.md#user-content-fn-51), [22](gemini-cli/article.md#user-content-fn-22)</sup> | Format/blocklist/syntax errors → silent requery (≤3) → `exit_format`; tenacity on API calls; error autosubmission diffs the tree into `/root/model.patch` - a failed run still submits something <sup>[10](swe-agent/article.md#user-content-fn-10)</sup> | `retryWithBackoff` (7 attempts, 1.5s→30s) + unattended infinite-retry mode; stream-transport retry allow-list + watchdogs; two-tier loop detector with result-aware guards; cross-restart session recovery with synthesized tool responses <sup>[89](qwen-code/article.md#user-content-fn-89), [90](qwen-code/article.md#user-content-fn-90), [91](qwen-code/article.md#user-content-fn-91), [92](qwen-code/article.md#user-content-fn-92)</sup> | Per-child 1.5h timeout + cancel/log; a generation never hard-fails; JSON extraction ×3; tool errors as strings; compile gate drops crashed children; no token/dollar accounting <sup>[3](dgm/article.md#user-content-fn-3), [13](dgm/article.md#user-content-fn-13), [17](dgm/article.md#user-content-fn-17), [18](dgm/article.md#user-content-fn-18)</sup> | Repair over termination: loop/reflection guards, deferred-action recoveries (≤2) and truncation repairs (≤3) all attempt in-loop fixes first; iteration exhaustion produces a summary result (`reason: 'iteration_limit'`) instead of a hard failure <sup>[13](autohand/article.md#user-content-fn-13), [14](autohand/article.md#user-content-fn-14), [15](autohand/article.md#user-content-fn-15), [11](autohand/article.md#user-content-fn-11)</sup> | Fail-closed throughout the tool path: an approval-answerer failure resolves `unavailable` → blocked; no enforcing sandbox backend → `SANDBOX_UNAVAILABLE` rather than an unconfined run; torn-tail log recovery on open; a provider-confirmed context-window error triggers a maximal-reduction compaction retry <sup>[24](deepseek-harness/article.md#user-content-fn-24), [25](deepseek-harness/article.md#user-content-fn-25), [35](deepseek-harness/article.md#user-content-fn-35), [17](deepseek-harness/article.md#user-content-fn-17)</sup> | Dangling tool calls left by interrupted runs are repaired on resume with synthetic error results; invalid assemblies fail at construction (protected-middleware exclusion, permissions+execution `NotImplementedError`, `task`-argument validation); when history offload fails, summarization degrades to an unrecoverable summary with a warning <sup>[14](deepagents/article.md#user-content-fn-14), [31](deepagents/article.md#user-content-fn-31), [48](deepagents/article.md#user-content-fn-48), [35](deepagents/article.md#user-content-fn-35)</sup> | A server-side doom-loop policy (the `x-grok-doom-loop-check` header) drives confident-signal resampling in the sampler; spilled upload queues are reconciled at startup — re-enqueued or purged by the current privacy verdict — and the coding-data opt-out itself defaults to opted-out <sup>[15](grok-build/article.md#user-content-fn-15), [36](grok-build/article.md#user-content-fn-36), [31](grok-build/article.md#user-content-fn-31)</sup> | Dead rollouts recorded, never raised; an invalid proposal gets one retry with the rejection fed back, then the stage inherits the current harness (fell_back); selector failure falls back to a delivered final; judge failure is None — excluded from the denominator, never a zero; resume is content-verified by harness fingerprint <sup>[6](harness2/article.md#user-content-fn-6), [13](harness2/article.md#user-content-fn-13), [26](harness2/article.md#user-content-fn-26)</sup> | Durable event delivery: per-agent dedup, batches acknowledged only after the turn completes, retry with backoff, stuck `sending` rows recovered on restart, session reloaded by id; bridge disconnects use native `session/cancel` with a settle wait; silent turns earn a forced-write nudge <sup>[11](agensh/article.md#user-content-fn-11), [24](agensh/article.md#user-content-fn-24), [27](agensh/article.md#user-content-fn-27), [43](agensh/article.md#user-content-fn-43)</sup> | Durable effects (retry, failure threshold 3); interrupted Assignments resumed with a recovery note; expired Evals requeued at most twice; sends idempotent by receipt + platform marker <sup>[44](dsh-superagent/article.md#user-content-fn-44), [52](dsh-superagent/article.md#user-content-fn-52), [40](dsh-superagent/article.md#user-content-fn-40)</sup> | Goal continuation is conservative: defers to background tasks, verifies before continuing, stops on verifier failure without a next action, re-reads goal state after verification so a user Stop wins; verifier JSON parsing fails open (recorded) rather than trapping a goal; compaction rapid-refill circuit breaker; turns are cancellable commands; provider configs fall back to last-known-good <sup>[19](zcode/article.md#user-content-fn-19), [20](zcode/article.md#user-content-fn-20), [4](zcode/article.md#user-content-fn-4)</sup> | Lease expiry (~10s TTL) and takeover by fold; fenced appends refuse stale holders in-transaction; unfinished calls rerun only when declared safe, else `interrupted`, with a stable idempotency key to MCP servers; poison guard seals turns whose workers keep dying; stall watchdog; cancel targets one turn by inbox position <sup>[10](everruns/article.md#user-content-fn-10), [11](everruns/article.md#user-content-fn-11), [34](everruns/article.md#user-content-fn-34)</sup> | Provider retry (3 attempts / 3 min, rejections never retried); failure limit with optional loop-breaker rewind; runs rebuild from rows after a crash; merge conflicts interrupt, never cancel; memory supervisor pauses the cheapest task <sup>[13](mike/article.md#user-content-fn-13), [21](mike/article.md#user-content-fn-21)</sup> |
| 10 | Security model | Project trust + extension hooks; no built-in approval UX <sup>[24](pi/article.md#user-content-fn-24), [33](pi/article.md#user-content-fn-33)</sup> | Chat-file edits auto-apply; prompts only at boundaries; git auto-commit + `/undo`; no sandbox <sup>[30](aider/article.md#user-content-fn-30)</sup> | Finest-grained: per-tool policies + webview UI + diff previews; commands prompt by default; no sandbox; per-turn stash checkpoints <sup>[3](cline/article.md#user-content-fn-3), [22](cline/article.md#user-content-fn-22), [40](cline/article.md#user-content-fn-40)</sup> | Modes × `permission.yaml` × 5-inspector chain (security can only escalate); LLM judge for SmartApprove (UNTRUSTED-labeled, fail-closed); Always-Allow withheld on security warning; strict headless; no undo <sup>[29](goose/article.md#user-content-fn-29), [30](goose/article.md#user-content-fn-30), [31](goose/article.md#user-content-fn-31), [32](goose/article.md#user-content-fn-32), [33](goose/article.md#user-content-fn-33), [28](goose/article.md#user-content-fn-28)</sup> | **Strongest: 4-layer sandbox** (Starlark exec-policy → `PermissionProfile` → OS backends: Seatbelt / bwrap+landlock+seccomp / Windows PSEC → MITM network proxy); **per-action approvals** (sealed enum); hooks → Guardian → user; `AskForApproval` 4-mode ladder; **no undo** <sup>[4](codex/article.md#user-content-fn-4), [15](codex/article.md#user-content-fn-15), [16](codex/article.md#user-content-fn-16), [17](codex/article.md#user-content-fn-17), [18](codex/article.md#user-content-fn-18), [21](codex/article.md#user-content-fn-21), [22](codex/article.md#user-content-fn-22)</sup> | Risk labels + confirmation policy + defense-in-depth rails; no OS sandbox in the loop <sup>[44](openhands/article.md#user-content-fn-44), [58](openhands/article.md#user-content-fn-58), [88](openhands/article.md#user-content-fn-88)</sup> | Permission gate per tool call: layered rulesets (session, approved, agent config), last-matching-rule-wins, default `ask`; `ask` parks on a `Deferred` until the user replies; answering `always` appends to the approved ruleset; no OS sandbox in the loop <sup>[16](opencode/article.md#user-content-fn-16), [17](opencode/article.md#user-content-fn-17)</sup> | Per-session gateway approval queue + YOLO-mode bypass + denial breaker; ACP edit approval (ContextVar-bound); subagent tool grants carved from parent, never widened; webhook payloads restricted to safe tool subset <sup>[18](hermes/article.md#user-content-fn-18), [27](hermes/article.md#user-content-fn-27), [19](hermes/article.md#user-content-fn-19)</sup> | Staged admission verdicts (ingest → classify → preflight → resolve → assemble → dispatch → finalize); explicitly ordered approval chain (ordering is behavior); use-time authority gate bound to the run's delegated authority; hashed-receipt audit; approval prompts as native channel UI (Telegram inline buttons) <sup>[4](openclaw/article.md#user-content-fn-4), [31](openclaw/article.md#user-content-fn-31), [32](openclaw/article.md#user-content-fn-32)</sup> | TOML policy engine (tiered priorities, first-match-wins) + confirmation bus; 4 approval modes; per-OS sandboxes (Seatbelt/bwrap/Windows tokens); Conseca LLM-written policy (opt-in, default off) <sup>[29](gemini-cli/article.md#user-content-fn-29), [34](gemini-cli/article.md#user-content-fn-34), [40](gemini-cli/article.md#user-content-fn-40), [39](gemini-cli/article.md#user-content-fn-39)</sup> | Sandboxing is the Docker container via SWE-ReX - agent runs as root inside; zero human approval prompts, blocklist only; hardcoded Flask secret - researcher threat model <sup>[20](swe-agent/article.md#user-content-fn-20)</sup> | AUTO default: rule-string `PermissionManager` (deny > ask > allow > default) + fail-closed LLM classifier + deterministic destructive-command guard + protected writes; confirmation bus gutted; Linux-only bwrap/landlock sandbox; YOLO settable in `settings.json` <sup>[36](qwen-code/article.md#user-content-fn-36), [43](qwen-code/article.md#user-content-fn-43), [47](qwen-code/article.md#user-content-fn-47), [37](qwen-code/article.md#user-content-fn-37)</sup> | Per-task Docker but unhardened — no network/memory/cap limits; zero approvals, no HITL; operator API keys as env vars; test-patch stripping is the sole hard anti-hack; full lineage traceability <sup>[19](dgm/article.md#user-content-fn-19)</sup> | Unbypassable security blacklist checked first (and re-checked); fail-closed dispositions — `default` collapses to prompt, unknown to deny; permission decisions stored as patterns; workspace trust keyed by a SHA-256 over the `{hooks, mcpServers}` config; extension runtimes are trusted in-process code — no sandbox <sup>[25](autohand/article.md#user-content-fn-25), [26](autohand/article.md#user-content-fn-26), [28](autohand/article.md#user-content-fn-28), [40](autohand/article.md#user-content-fn-40)</sup> | Two-knob permission presets (sandbox mode × approval policy: read-only / workspace-write / danger-full-access) pinned per session in the log; per-call OS sandboxing — bubblewrap → Landlock, Seatbelt, Windows ACL (reported `partial`) — with a one-call approval-gated escalation ladder; **plugins run in-process outside the sandbox**; session-log upload to the official API is default-on (switchable) <sup>[23](deepseek-harness/article.md#user-content-fn-23), [26](deepseek-harness/article.md#user-content-fn-26), [29](deepseek-harness/article.md#user-content-fn-29), [30](deepseek-harness/article.md#user-content-fn-30)</sup> | The backend is the boundary: the default `StateBackend` cannot execute at all; `LocalShellBackend` runs host commands via `shell=True`, unsandboxed, under extensive docstring warnings; real isolation only via `BaseSandbox` partner backends; first-match glob permission rules (allow/deny/interrupt) fence the fs tools — symlink-resolved, listings filtered — but do not cover `execute` <sup>[28](deepagents/article.md#user-content-fn-28), [29](deepagents/article.md#user-content-fn-29), [24](deepagents/article.md#user-content-fn-24), [31](deepagents/article.md#user-content-fn-31)</sup> | Six permission modes; authorization runs hooks → rules (deny > ask > allow) → per-project grants → read-only auto-approvals → mode policy, with `auto` mode decided by an LLM transcript classifier; an opt-in whole-process OS sandbox (Landlock/Seatbelt) is applied at startup before the agent boots, off by default; whole-repo collection is stubbed out at this pin — both producers return 'unavailable in this build' <sup>[19](grok-build/article.md#user-content-fn-19), [20](grok-build/article.md#user-content-fn-20), [21](grok-build/article.md#user-content-fn-21), [29](grok-build/article.md#user-content-fn-29)</sup> | Score-blindness by construction (RunView carries no score) plus tripwires: grader-verdict marker scans on every evidence/harness/library tree, per-benchmark answer-key leak globs, anonymized selection; proposer sessions in fenced temp workspaces with API keys stripped from the environment; the executor side is open by design — network unrestricted, Codex approvals/sandbox bypassed, mitigation delegated to operator egress filtering and disclosed in the docs <sup>[15](harness2/article.md#user-content-fn-15), [27](harness2/article.md#user-content-fn-27), [53](harness2/article.md#user-content-fn-53)</sup> | No runtime approvals (`--no-ask-user`) and zero-approval self-merge by design; guardrails are subtractive (60-entry MCP deny list) and credential-scoped (per-worker tokens from an auth-gated secret service, board membership checks, workers as UID/GID 10000); offline worker networks and an execute-only reference binary in the benchmark condition <sup>[21](agensh/article.md#user-content-fn-21), [38](agensh/article.md#user-content-fn-38), [46](agensh/article.md#user-content-fn-46), [48](agensh/article.md#user-content-fn-48)</sup> | Admission gate + assessor judgement up front and a Mandate per Ticket; after accept, workers run without approval as the operator's OS user at `danger-full-access` — blocked commands are PATH shims, containment is the dedicated account <sup>[34](dsh-superagent/article.md#user-content-fn-34), [30](dsh-superagent/article.md#user-content-fn-30), [66](dsh-superagent/article.md#user-content-fn-66)</sup> | Six collaboration modes (plan/build/edit/yolo/guarded/auto) with a per-tool permission service (risk, side-effect scope, rulesets, plan-mode policy) and hook-visible PermissionRequest events; containment rests on that layer plus bash command analysis — no OS sandbox in the core at this pin; bots bound to a single provider user id with workspace/command allow-lists and credential-service secrets; workspace checkpoints are local-only at this pin, telemetry endpoints build-gated (empty = off), OTLP traces environment-gated <sup>[11](zcode/article.md#user-content-fn-11), [22](zcode/article.md#user-content-fn-22), [33](zcode/article.md#user-content-fn-33), [38](zcode/article.md#user-content-fn-38), [41](zcode/article.md#user-content-fn-41)</sup> | Org-scoped multitenancy at the API layer; agent keys scoped to one channel's own sessions; hard approval gate keyed on declared tool risk (normal/protective/off), fingerprint-bound one-off approvals answered against the emitted event; scoped write-only credentials; envelope-encrypted secrets; egress policy; workers hold no database credentials <sup>[16](everruns/article.md#user-content-fn-16), [37](everruns/article.md#user-content-fn-37), [42](everruns/article.md#user-content-fn-42)</sup> | gVisor (runsc) per task via rootless Podman, `--cap-drop ALL`, no-new-privileges; worktree rw with its `.git` pointer pinned read-only, project `.git` read-only; `--network none` an explicit per-project choice; missing bridge refuses tools; unsandboxed needs `DANGER_ZONE=1`; passkey pairing <sup>[15](mike/article.md#user-content-fn-15), [17](mike/article.md#user-content-fn-17), [48](mike/article.md#user-content-fn-48)</sup> |

Superscript numbers link to the evidence endnote in that column's article.

## Columns

- **pi (v1.0.2)** — [article](pi/article.md) · source: [earendil-works/pi](https://github.com/earendil-works/pi)
- **aider (v0.86.2)** — [article](aider/article.md) · source: [Aider-AI/aider](https://github.com/Aider-AI/aider)
- **Cline (v4.1.22)** — [article](cline/article.md) · source: [cline/cline](https://github.com/cline/cline)
- **Goose (v1.53.0)** — [article](goose/article.md) · source: [block/goose](https://github.com/block/goose)
- **Codex CLI (rust-v0.160.0)** — [article](codex/article.md) · source: [openai/codex](https://github.com/openai/codex)
- **OpenHands (v1.53.0)** — [article](openhands/article.md) · source: [OpenHands/software-agent-sdk](https://github.com/OpenHands/software-agent-sdk)
- **OpenCode (v1.18.34)** — [article](opencode/article.md) · source: [anomalyco/opencode](https://github.com/anomalyco/opencode)
- **Hermes (v2026.9.24)** — [article](hermes/article.md) · source: [NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent)
- **OpenClaw (v2026.9.9)** — [article](openclaw/article.md) · source: [openclaw/openclaw](https://github.com/openclaw/openclaw)

- **Gemini CLI (v0.63.0)** — [article](gemini-cli/article.md) · source: [google-gemini/gemini-cli](https://github.com/google-gemini/gemini-cli)
- **SWE-agent (v1.1.0)** — [article](swe-agent/article.md) · source: [SWE-agent/SWE-agent](https://github.com/SWE-agent/SWE-agent)
- **Qwen Code (v0.25.0)** — [article](qwen-code/article.md) · source: [QwenLM/qwen-code](https://github.com/QwenLM/qwen-code)
- **DGM (a565fd2)** — [article](dgm/article.md) · source: [jennyzzt/dgm](https://github.com/jennyzzt/dgm)
- **Autohand Code CLI (v0.9.8)** — [article](autohand/article.md) · source: [autohandai/code-cli](https://github.com/autohandai/code-cli)
- **DeepSeek Harness (dsh-v0.2.1-alpha.3)** — [article](deepseek-harness/article.md) · source: [deepseek-ai/deepseek-harness](https://github.com/deepseek-ai/deepseek-harness)
- **Deep Agents (v0.7.25)** — [article](deepagents/article.md) · source: [langchain-ai/deepagents](https://github.com/langchain-ai/deepagents)
- **Grok Build (2bdd1d6)** — [article](grok-build/article.md) · source: [xai-org/grok-build](https://github.com/xai-org/grok-build)
- **Harness² (1cfe254)** — [article](harness2/article.md) · source: [google-research/harness2](https://github.com/google-research/harness2)
- **Agensh (v1.0)** — [article](agensh/article.md) · source: [microsoft/Agensh](https://github.com/microsoft/Agensh)
- **dsh-superagent (v0.1.0)** — [article](dsh-superagent/article.md) · source: [deepseek-ai/dsh-superagent-prototype](https://github.com/deepseek-ai/dsh-superagent-prototype)
- **ZCode (3.16.2 @ 3c5b227)** — [article](zcode/article.md) · source: [zai-org/ZCode](https://github.com/zai-org/ZCode)
- **Everruns (v0.49.0 @ 4e8002a)** — [article](everruns/article.md) · source: [everruns/everruns](https://github.com/everruns/everruns)
- **MIKE (v0.1.0 @ 382ec61)** — [article](mike/article.md) · source: [SoftwareRenderer/MIKE](https://github.com/SoftwareRenderer/MIKE)
