GuidesAI and agents

Orchestrating AI sub‑agents: splitting a task across several agents

By the Banzai Labs teamUpdated on 2 min read

OrchestratorResearchAnalysisWritingSynthesisEach sub-agent has its own context; only the summary comes back
An orchestrator hands tasks to sub-agents, each with its own context, then assembles their summaries.

The pattern

  1. 1PlanThe orchestrator splits the task.
  2. 2DelegateOne precise brief per sub-agent.
  3. 3WorkIn parallel, each in its own context.
  4. 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

The right choice depends on the task
A single agent
  • Short, sequential task
  • Fewer calls, lower cost
  • Easier to debug
Orchestrator and sub-agents
  • Independent parts, in parallel
  • Light main context
  • More calls to monitor

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.

Read next

All guides