Part 9 of 11 · 1 min read
Quick notes
- Start with one agent. It is simpler, cheaper and faster; add agents only when the task, the number of tools or the need for coordination outgrows one.
- Three reasons justify splitting: parallelism of independent sub-tasks, context or tool overload, and specialisation.
- The reasons against: every sub-agent costs a system prompt, a copy of the shared background, a brief and a result, and each handoff between dependent sub-tasks can lose a fact. A multi-agent system uses many times the tokens of a single chat.
- The deciding quantities are the length of the dependency chain, the number of facts that must cross between sub-tasks, the size of the shared background, and whether a prompt cache already makes one long context cheap.
- Split by independence, not by job title. In the model, a full chain of dependent sub-tasks was slower than one agent, and a prompt cache plus a large shared background flipped the token saving into a loss.