When a parent agent hands off to a sub-agent, what is the minimum context it should pass and why?

A parent agent often needs help from a more specialized sub-agent: a search specialist, a code reviewer, a planner. The question is what context to send across that boundary. The answer is a **minimal, self-contained brief**. That brief includes four things: the sub-task statement (one or two sentences saying exactly what the sub-agent must do), the inputs (the data the sub-agent will work on), the output shape (what structure the parent expects back), and any constraints that bind the sub-task (latency budget, tone, refusal rules). Deliberately leave out the parent's full scratchpad, the parent's prior conversation turns, and tool results that are not relevant to this sub-task. Those things are noise from the sub-agent's perspective. Extra context lowers precision, inflates cost, and worst of all, can mislead the sub-agent into answering the parent's original question instead of the narrow job it was given. The mental model is a function call. You pass arguments and you get a return value. You do not hand the callee a copy of the caller's stack frame.

**A minimal, self-contained brief.** Sub-task statement, the inputs the sub-agent needs, the expected output shape, and any binding constraints. Nothing else. The two reasons matter equally. First, narrower work needs narrower context. The sub-agent's job is more constrained than the parent's; loading it with parent-scope information dilutes the signal to noise ratio of its prompt and degrades precision on the actual sub-task. Second, leakage causes scope creep: a sub-agent that sees the parent's whole conversation tends to drift toward solving the parent's overall problem rather than the bounded slice it was handed. The handoff stops being a contract and becomes a fork. The contract framing makes this concrete. A handoff has a signature like a function: `subagent(task, inputs, output_schema, constraints) -> structured_result`. Everything outside that signature is leakage. The parent integrates the structured result back into its own state; the sub-agent never sees the integration logic. A practical heuristic: if you cannot write the brief in under 500 tokens, the sub-task is either too broad to delegate or not actually separable from the parent. Re-decompose before spawning.