Multiple agents: when splitting work helps and when it gets in the way
Use independent tasks, verifiable deliverables and accountable integration to test whether multiple agents improve the outcome.
Three agents edit the same component. Each solves part of the task, and the final integration undoes half the changes. There is plenty of activity. The division created dependencies without deciding who owns the combined result.
Before adding agents, ask what can progress independently. Investigating separate topics differs from editing the same authorisation rule in parallel. The number of participants should follow the structure of the work.
Find boundaries with verifiable deliverables
In a hypothetical product review, one stream examines navigation, another API contracts and another copy. Each needs an objective, the necessary materials and a clear description of what to return.
Avoid assignments such as “improve quality”. Request a bounded result: reproducible findings, changes to named files or a source comparison with identified passages. The integrator should be able to assess it without repeating the entire investigation.
Anthropic describes a research system with coordination and parallel subagents. It is an account of one product, not proof of improvement for every task.
Decide who may change what
For code, separate files or use isolated checkouts where changes must coexist. If two results alter the same contract, return that decision to coordination before merging. A “done” message does not resolve conflicting behaviour.
For research, ask each stream to return sources and limitations. Agreement between agents reading the same summary is not independent confirmation. Whoever assembles the report must check the claims and the connection between evidence and conclusion.
Count the coordination work
Compare time to acceptance, total cost, conflicts and review effort. A parallel stage may finish sooner and still require enough integration work to delay the whole project.
Try the same kind of task with a simple execution and a divided one. Record where division helped and where it increased rework. File counts and message volume are poor substitutes for accepted outcomes.
Assign responsibility for the final result
Someone must combine deliverables, resolve disagreements and validate the complete workflow. A subagent can prove its part while the combination remains wrong. Include tests that cross task boundaries.
Keep small or sequential work simple. Use multiple agents when there is useful parallelism, capacity to review the results and an observed benefit in the final outcome.
Continue with preparing a change for review.
About the author
Tiago F SantiagoComments
No comments yet
Share a question or an experience related to the article.
Related articles
