How I Built Observability for My Autonomous Coding Agent: 5 Lessons

Short summary

A developer shares five lessons from building an observability layer for an autonomous coding agent that runs unattended for hours. The key insight: logging must sit in the harness between agent and tools, not inside the agent's own reasoning, because the agent's self-narration can be wrong. Distinguishing 'no-op' tool results from 'success' proved critical for catching silent partial failures where the agent claimed work it didn't actually do.

• •Log tool calls from the harness, not the agent — agent self-narration can be as wrong as its summaries
• •Track 'no-op' status separately from 'success' to catch silently skipped edits and partial failures
• •JSONL append-only logs are crash-safe, greppable, and require no database

Generated with AI, which can make mistakes.