What is worth delegating

A subagent is a separate Claude with its own context, its own instructions and its own tool list. It gets the brief you hand it — not your conversation — and hands back a result. That isolation is the entire value and the entire constraint.

It's valuable because a long investigation can happen without filling your session with the eighty files it read. It's constraining because everything the worker needs must be in the brief. So the test for a delegable task is: can I state the inputs, the question and the output format without reference to our conversation? If yes, delegate. If the task only makes sense given three hours of shared context, doing it yourself is faster than explaining it.

The brief that follows should contain the goal, the paths that matter, the constraints, and the shape of the answer — and nothing else. Overlong delegations are the common beginner error: pasting the whole conversation in the hope that more context helps. It doesn't. It buries the actual question and spends the worker's context on your history rather than on the files it needs to read.

The other judgement call is when not to parallelise. Independent questions run happily side by side. Dependent steps do not: if step two needs step one's answer, running them together means step two guessed. Map the dependencies before you fan out, and keep the dependent chain sequential — parallelism is a property of the work, not a setting you turn on.

Loading your workspace…