Agent Delegation
Agent delegation is the pattern where one agent hands a scoped sub-task to another, with inputs, success criteria, and an expected return shape.
In plain English
Delegation in a multi-agent system is not just "call this other agent." A good delegation packs four things: the goal, the inputs, the definition of done, and the shape of the return value. The called agent gets enough context to act but not so much that it drowns. The caller gets a predictable result it can plug into the next step.
Framed this way, delegation is just function calling, with a twist — the "function" is another agent with its own reasoning loop. The contract still matters. If the caller says "return a headline and three bullets," the callee must honor that shape. When the shape is loose, downstream steps start guessing, and the whole orchestration becomes fragile.
Why it matters for Black Box
The CEO agent delegates to specialists hundreds of times per task. Each delegation includes a goal, inputs, acceptance criteria, and a return shape — often enforced by structured output. That discipline is what lets Black Box chain twenty agent turns without the output drifting.
Examples
- CEO agent delegates "draft the hero copy" to the Content agent with brand voice and length constraints.
- CEO agent delegates "review this PR" to the Evaluator with a checklist and a pass/fail return shape.
- Content agent delegates "fetch competitor pricing" to the Browser agent with a target URL and a CSV output spec.