Delegation Basics

You hand one model a goal with nine moving parts — pull the numbers, check them against last quarter, draft the summary, file the ticket — and it does four, forgets two, and invents the rest. Nothing was broken. You gave one context window a job that needed a team.

An agent is a model that can take actions, and delegation is the first pattern you reach for when one agent stops coping. A planner reads the goal, splits it into subtasks, and routes each one. Workers execute, each holding only the slice of context its own subtask needs and only the tools that subtask requires. The planner never runs the work itself.

The value is in the hand-off, so write it down. A worker gets three things: the task, the inputs it needs, and a definition of done someone else could check. It returns two: the result, and what it could not do — the source it never found, the field left blank. The planner decides what gets done and who does it. How it gets done is the worker's call.

That second value is the one people drop, and it is what makes the pattern recoverable. A worker that reports a gap lets the planner re-route. A worker that fills the gap with a plausible guess hands you a finished-looking document with a hole in it.