Lesson 03 · Orchestration · 12 min

Sub-agents

The maker and checker in the lab are a sub-agent pattern: one job, two agents, separate contexts. Generalising that is how loop engineering handles work too large or too varied for a single loop.

There are two distinct reasons to split, and they are worth keeping apart because they suggest different designs.

IndependenceYou want a judgement not contaminated by the work. Maker and checker. Implementer and reviewer. The value is the separation itself.

ContextThe job needs more material than one context can hold, or the material for one part would distract from another. The value is what each agent does not have to know.

The second is underrated. An agent given every tool and every document reasons worse than one given the four tools its job requires — not because the context is full, but because relevance gets harder to establish as irrelevance grows.

Give each sub-agent the narrowest set of tools that lets it finish.
Pass conclusions between agents, not transcripts. A summary is a contract; a transcript is a leak.
Keep the verifier ignorant of how the work was produced, or it will start agreeing with the reasoning.
Be honest about cost: two agents are at least two calls, and the split needs to earn that.

Multi-agent systems are frequently reached for too early. If one agent with a good harness handles the task, the second agent adds latency, cost and a coordination problem. Split when there is a reason above, not because a diagram with more boxes looks more serious.

Checkpoint

Name a job in your own work that would benefit from an independence split, and one from a context split.