Autonomous Agent Loop Patterns — What Actually Works

After building and testing multi-model agent loops (Opus/Sonnet/Codex), the patterns that produce reliable output vs ones that spiral are becoming clear. The core working pattern is: plan → execute one atomic task → verify output → decide next step. Loops that try to plan 10 steps ahead and execute sequentially fail because error compounds — step 3 assumptions break by step 6.

The verification step is non-negotiable. Without it, agents confidently produce broken code, hallucinated data, or circular rewrites. Effective verification means: run the code, check the output matches spec, diff against previous version. Human-in-the-loop works best as a gate between phases, not between individual steps — review after "build the feature" not after every file edit.

Key insight from our system: the orchestrator (ClawdBot) should own project state and task queue, delegating heavy work to specialized agents (Codex for code, Opus for architecture, Sonnet for quick tasks). The orchestrator never does the work itself — it assigns, monitors, and routes. Agent verbosity is a real capability signal: an agent that can't calibrate response length to context wastes attention budget even when other friction is solved.