Going deeper

Two published positions

In June 2025, two engineering teams published opposite advice within days. Cognition’s Don’t Build Multi-Agents argues that parallel agents make conflicting decisions, because each one sees only part of the context. Anthropic’s account of its research system shows an orchestrator and parallel workers doing better on broad research, while using about 15 times the tokens of a chat.

They aren’t really in conflict. Research splits into independent questions, and reading more sources in parallel helps. Writing code doesn’t split that way: each decision constrains the next. So ask whether subtasks are independent. If a worker needs another worker’s choices to do its job, keep the work in one agent.

Patterns

• Supervisor: one agent delegates and combines. Easy to follow, and it becomes a bottleneck.
• Handoff: control passes from one agent to the next, as in a support desk. There is one speaker at a time, and context can be lost at each transfer.
• Blackboard: agents read and write a shared state. It’s flexible, and it needs rules about who may write what.

Share decisions, not transcripts

Forwarding a whole conversation wastes tokens and buries what matters. Forwarding one sentence loses the reasons. A small structured state works better:

{
  "goal": "Refund order 4512",
  "decisions": ["Damaged item: policy P-7 applies"],
  "open_questions": ["Was the item already replaced?"],
  "artifacts": ["orders/4512.json"]
}

A vague brief is the commonest failure. “Look into the refund” leads a worker to redo the supervisor’s research. A good brief gives the goal, what’s already known, the expected format and when to stop.

Fan-out and fan-in

Fan-out needs a concurrency limit, because rate limits are shared. It also needs a decision for partial failure: if one worker in five fails, is the result usable? Fan-in is where the risks gather. Workers can disagree, repeat each other or return made-up facts, so the merge step needs its own instructions and its own tests.

Errors multiply

Five steps that each succeed 95% of the time succeed together about 77% of the time (0.95⁵). Each agent you add is another factor in that product, unless a later step actually catches errors. A reviewer agent only helps if you’ve measured that it finds them.