Multi-agent task routing: how Organ's agents decide who does what

说明与用途

Organ 的多代理路由器通过 SQL 证据、路由代理、审查代理和代码不变式来决定工作类型、所有者和时间，消除了工具绕过、缺失审查和重复工作，将完成率提升至 86.3%。

问题

Agents previously chose workflow types and tools, leading to routing around typed tools, bypassing reviews, lost tasks, and duplicate work.

结果

The router handled 1,235 requests, completing 909 runs, denying or combining 155, and failing 169. Completion rate 86.3%. Model token cost $708.50 across 1,113 runs, averaging at least $0.77 per started run. The system now records failure causes, separates dedup keys, and tracks denial rates.

工作流

• Gather evidence in code (SQL)
• Ask routing agent for shape, owner, timing
• Review decision against policy in a separate turn
• Apply invariants in code
• Dispatch or deny with a way forward

证据

Agents routed around the typed tools. A deprecated catch‑all tool, labelled "do not use" in its own description, took 457 calls. All five typed tools together took 232.

来源

The type an agent picked could skip review. 32 of 58 OPS generic-task runs, and 18 of 41 from my own marketing department, were really code changes. Generic tasks have no review or validation phase, so that code shipped without one.

来源

The router also doesn't start its own container. Starting one averaged 775 seconds (p90 1,586) when we designed this, so the router runs inside the requesting agent's container, which already exists.

来源

The main problem now isn't that routing breaks. It's that we can't see why it breaks.

来源