One agent can’t be an expert at everything.
Build a team of specialists, each with its own instructions, tools and model: a reviewer that only reviews, a test writer that only writes tests, a researcher that never edits a file. The lead reads the task, picks the right specialist and hands the work over.
Watch the lead delegate as it happens: who is working on what, how much it took, and what came back. Open any specialist to see exactly what it did.
Save an agent to the project and it lives in the repo, so everyone works with the same specialists. Keep personal ones global and use them everywhere.
Stop prompting one generalist. Start running a team.
An agent needs a name, a line saying when to use it, and instructions. The lead reads that line to decide when to pass work along, so it's worth writing well. Then pick its model and the tools it's allowed to use.
| Name | One word, like security-reviewer. You and the lead call it @security-reviewer |
| Use when | "Reviews a change for injection, auth and secret-handling bugs. Use after edits that touch input parsing or credentials." |
| Instructions | What it does, how it should work, and what to report back |
| Context | Start from the task alone, or pick up the conversation so far |
| Model | Same as the session, or its own from any provider you've connected |
| Tools | Everything the lead can use, or just a few things |