Notes on working with AI
Give the lead agent no hands
Told to use subagents when appropriate, a model delegated nothing in a reported test. Taking away its tools worked where wording didn't.
Five runs, zero delegations
Suppose you run several agents at once. The obvious design is one chat in charge that hands work to subagents, with a line in its instructions like "use a subagent for a piece of work when appropriate." DeepSeek reported testing that line: a model, about 20 real work requests mixed into roughly 300 chat and CI messages, five runs. It never called a subagent. It did the work itself, and every new message interrupted it.
Telling it harder helps, until it doesn't
A firmer sentence did better: every message reaches you, so work you do yourself keeps stopping for new messages; hand each piece of work to a subagent and do only small things yourself. That delegated in four of five runs. In the fifth, the model judged the job small and did it itself. The word "small" was a door in the prompt, and on the fifth run it walked through. Delete the word and the model still decides what counts as a piece of work.
Taking the hands off
DeepSeek's fix wasn't a better sentence. The lead gets no coding tools at all. It can only receive tasks, start workers, and relay results. Each task gets its own worker, so the mess stays with that worker, and the CI log stays with the fix. In the reported test the lead without coding tools delegated every time. The numbers come from DeepSeek's own small experiment, one setup per condition, so read them as a signal, not a ranking.
Why it matters to you
With one task you wait for the model. With many you wait for yourself: a person can usually follow three to five sessions. One chat per task turns you into the courier who copies paragraphs between tabs. A lead that only coordinates gives you one place to talk.
Three ways to try it
Claude Code. In the instructions for your coordinator session, write that it never edits files or runs commands itself, and only delegates to subagents and reports.
Finance close. A Claude Project whose job is to split month-end into reconciliations, accruals and flux commentary, and write a complete brief for each. You run each brief in its own chat and bring the results back.
Marketing. A weekly campaign: research, draft and review each get their own brief from the coordinator.
The rule underneath all three: don't leave the lead a task it can decide is small.