XiaoyaoClaw Agent Orchestrator
Multi-agent orchestrator — wraps "split → dispatch → track → aggregate → retry" into a standard workflow so persistent agents collaborate like a team.
Product Introduction
OpenClaw Agent Orchestrator tackles five pain points of multi-agent parallel work: two-layer config with bidirectional allow, fragile manual channels, 60s timeouts (idle? dead?), scattered replies needing manual aggregation, and failure-stall cascades. Solution: "one-liner dispatch → auto fan-out → transparent progress → sourced aggregation → auto retry". Tasks go to peer’s persistent sessions (full persona + memory + skills — like real colleagues). Three trigger modes: direct on name/verb, ask-once on fuzzy big tasks, silent on casual Q&A. Multi sub-task fan-out (fire-and-forget), wall time ≈ slowest. Live status via sessions_list / sessions_history — 60s timeout ≠ failure. Aggregation labels "which conclusion from which agent". Default 3 retries with context (no redoing completed work). Reads openclaw.json bidirectional allowlist directly; on missing config, ask first, then safe config.patch merge. Config + status scripts are pure Python stdlib, Windows / macOS.
Core Features
- Forced sessions_send — tasks go to peer’s persistent session (persona + memory + skills)
- Three triggers — name/verb → execute; fuzzy big task → ask once; casual → stay silent
- Parallel fan-out — fire-and-forget; wall time ≈ slowest
- Transparent progress — sessions_list/history live; 60s timeout ≠ failure
- Sourced aggregation — labels "which conclusion from which agent", traceable
- Retry — default 3 (configurable), context-aware, escalate after 3
- Reads openclaw.json directly — agents.list + agentToAgent.allow, single source of truth
- Zero-dep scripts — config detect + status aggregation, pure Python stdlib, cross-platform