Agents
Multi-agent system
A multi-agent system splits work across several agents with different roles, tools or models, passing control between them.
It earns its cost when sub-tasks genuinely need different tools or different models. Splitting merely to give each agent a tidier prompt rarely pays.
Every agent carries its own fixed overhead on every request it handles, so a six-agent design pays six system prompts and six tool schemas before any work happens.
In practice
Every handoff between agents is a context transfer, and the token cost of restating what the previous agent knows is the part people forget when budgeting. Multiple agents earn their overhead when the subtasks genuinely need different tools or permissions; they do not when the split is organisational rather than technical.
Common questions
When is a multi-agent system worth it?
When subtasks genuinely need different tools or different permissions. When the split is organisational rather than technical, you have paid handoff costs for an architecture diagram.
What is the hidden cost of multiple agents?
Every handoff restates context. The tokens spent telling the next agent what the previous one learned are real and are usually omitted from the budget that justified the split.
Measure it
- Multi-Agent Handoff Prompt BuilderStructured handoffs that stop agents repeating each other’s work.
- Subagent Token Budget PlannerBudget allocation across subagents, with fixed overhead subtracted first.
- Agent Loop Cost SimulatorPer-iteration breakdown showing why agents cost several times the naive estimate.