Agent Loop (OpenClaw)

The agent loop is the full execution path in OpenClaw: intake → context assembly → model inference → tool execution → streaming replies → persistence.

See Source - Agent Loop (OpenClaw) for the detailed source summary.

Quick reference

• Fire-and-forget: agent RPC → returns {runId, acceptedAt} immediately.
• Wait for completion: agent.wait RPC → blocks until lifecycle end/error.
• Runs are serialized per session via session lanes (+ optional global lane).
• Timeout hierarchy: agent.wait (30s default) < LLM idle (60s default) < agent runtime (48h default).

Hook entry points (ordered)

1. before_model_resolve — override model
2. agent:bootstrap — adjust bootstrap context files
3. before_prompt_build — inject context/system-prompt additions
4. before_agent_start — legacy
5. before_agent_reply — synthetic reply or silence
6. before_tool_call / after_tool_call — intercept tools
7. tool_result_persist — transform results before transcript write
8. agent_end — post-completion inspection
9. before_compaction / after_compaction — compaction cycle

Related

• OpenClaw — system overview
• OpenClaw Hooks — hook reference
• Command Queue — lane/queue modes
• Streaming — event stream details
• Compaction — auto-compaction behavior