Embedded Harness Anatomy
Top-down dissections of agent harnesses that run on milliwatts. Same discipline as the main series — pinned source, every claim verifiable — adapted for event-driven systems under physical constraints: power budgets, KB-scale memory, no guaranteed connectivity.
E1 — ESP-Claw: the agent loop that waits for an invitation
A reactive embedded system first and an agent second. Every input — an instant message, a schedule tick, a boot event — lands on an event bus; router rules decide what happens next, and the fast paths complete in milliseconds with no LLM involved. The LLM sits in the cloud, as this hardware dictates; on-device, safety is architectural rather than gated, and Lua is not sandboxed. Espressif's reference design for agents on the ESP32.
E2 — MimiClaw: the turn loop that moved onto the chip
The inverse of E1: not a bridge to a computer running the agent, but the whole OpenClaw-style loop reimplemented in pure C on a $5 ESP32-S3. Every input — a Telegram message, a cron firing, a heartbeat tick — becomes a cloud LLM turn; there is no local fast path, no approval gate, and memory is plain files on flash that the model edits itself. A single-consumer queue and a 21-pin GPIO allowlist are the whole safety story at v0.1.1.
E3 — ZenClaw: persistence that routes around the flash
The series' first Rust harness (esp-idf-svc, ESP32-S3 and P4) and its first real answer to flash wear: hot state lives in a PSRAM cache, every write replicates write-through to an S3-compatible bucket — a coalesced eager path for sessions, a blocking strict path for memory, cron, and config — and flash receives one atomic snapshot every 15 minutes. Slash commands are the only local fast path; the HTTP API is unauthenticated by documented design; cron and heartbeat are scaffolding the README itself labels unwired.