Agent Orchestration
When to dispatch one agent versus a swarm, how to write a brief that does not drift, and the shape of an agent's exit state.
Principle
The best results come from multiple agents working in sequence, each with a narrow scope and a clear handoff.
Techniques
1. Decompose Before You Dispatch
Break the task into subtasks before assigning agents. A vague brief to a capable agent produces plausible guesses. A specific brief produces execution.
2. One Agent Per Craft
Do not ask one agent to research, write, and edit. Dispatch a researcher, then a writer, then an editor. Each agent holds one role.
3. The Bible Reads First
Every persona should read its context before it speaks. The first message in any agent thread should surface the relevant background.
4. Render at the Project Level
Plugins ship prompts. Projects ship renders. The API keys, idempotency state, and filesystem layout live in the project — not the plugin.
5. Monitor the Seams
The handoff between agents is where quality lives or dies. Review the output of each stage before feeding it to the next.
When to Use
Any complex creative or engineering task where one pass is insufficient: essays, videos, software features, research reports.
Further reading
- "The Orchestrator Does Not Write" — why the persona that decides who writes is structurally separate from the personas that write.
- "The Hand-off" — the psychology of trust when one agent's output becomes another's input.
- "Personas Are Channels" — why something other than the persona has to decide when to call them.