GuidesAI and agents
Orchestrating AI sub‑agents: splitting a task across several agents
The pattern
- 1PlanThe orchestrator splits the task.
- 2DelegateOne precise brief per sub-agent.
- 3WorkIn parallel, each in its own context.
- 4AssembleThe orchestrator reads the summaries and concludes.
A good delegation brief
| Element | Example |
|---|---|
| Goal | “List the termination clauses in this contract” |
| Scope | “Only the attached document” |
| Tools | Read only, no sending |
| Return format | JSON list, 200 words max |
| Limit | 8 steps |
One agent or several
A single agent
Orchestrator and sub-agents
The pitfalls
- Vague briefs: each sub-agent redoes the others’ work.
- Returns that are too long: the orchestrator’s context overflows anyway.
- No limit per sub-agent: costs add up quickly.
Before this: the AI agent loop. The starting choice: AI agent or automated workflow.
Frequently asked questions
- When do you need several agents?
- When the task splits into independent parts (searching several sources, analyzing several documents) or when a step requires a lot of reading of which you only want the conclusion.
- Is it more expensive?
- Often, yes: each sub-agent makes its own calls. The gain is in quality and speed, not cost. Start with a single agent.
- Do sub-agents talk to each other?
- In the simplest pattern, no: they report to the orchestrator, which decides what comes next. It is easier to follow and debug.