Find the work that is actually independent
Start by running the task with one agent. Record time, cost, failure cases and the quality of the result. Add parallel work only where one worker can produce something useful without waiting for another. Different public-source searches may be independent; editing the same file usually is not.
A team creates coordination work: instructions must travel, artifacts must be reconciled and mistakes can propagate through summaries. More agents can widen coverage, but also repeat the same assumption. Name the particular benefit you expect from each added role before giving it a runtime.
Make every handoff a small contract
For a research brief, a source collector finds passages, an analyst compares claims and a reviewer attacks the final draft. Give each role a bounded input, an output format, ownership and a stop condition. Pass source references and artifacts rather than a growing pile of conversational instructions.
Assign one final owner to resolve disagreements. If two workers edit separate files, specify who integrates them and which checks establish compatibility. A shared workspace is a coordination mechanism, not permission to overwrite another worker's changes. Retain contradictory evidence until someone resolves it explicitly.
Protect the value of the reviewer
The reviewer needs enough independence to notice what the maker missed. Give it the original task, exact output and acceptance criteria. Ask it to locate a counterexample, not merely to approve. A different provider may reveal different blind spots, but still needs evidence and can still be wrong.
Compare the team with your single-agent baseline on the same held-out cases. Measure coordination failures as well as answer quality. Keep the team only if its benefit justifies the additional latency, cost and operational complexity. Delegation is a design choice to test, not a badge of sophistication.
Explore a bounded agent loop
The same workflow. Three different outcomes.
- 1Observe
- 2Prepare
- 3Check
- 4Done
Observe
Three public source files are available. The requested output is a local comparison draft.
An illustrative trace. No model, tool or external action is executed.
A worked example · Illustrative
A three-role research team
Collector → evidence.json
Owns: primary-source passages and dates
Analyst → comparison.md
Owns: synthesis; every claim links to evidence
Reviewer → review.md
Owns: counterexamples and unsupported-claim checks
Integrator: human project owner
Shared rule: no publishing tool; no overlapping file ownership
Stop: unresolved source conflict or exhausted budgetPut the idea to work
Your practice
- Run and save a single-agent baseline on a public sample.
- Split one independent responsibility into a second role.
- Define exact outputs, ownership and failure handling for each handoff.
- Have the reviewer challenge a hidden flawed claim and compare outcomes.