Guide · Agent orchestration
When One AI Agent Is Enough
Decide whether a task needs multiple agents by looking at dependencies, context, and the cost of combining their work.
Before you build a team of agents, give one agent a fair attempt at the task. Use a clear brief, the necessary tools, and an example of an acceptable result. If it finishes reliably, you already have a useful baseline. More workers should solve a specific problem that you can name.
Look for work that can stand alone
Reviewing three unrelated supplier documents offers a possible split. Each worker can extract the same fields from one document, and a coordinator can combine the findings. Rewriting three paragraphs of the same short email is a weaker candidate because every paragraph depends on the same voice, offer, and argument.
Anthropic's workflow guidance distinguishes fixed sequences, routing, parallel work, and more flexible agent behavior. Treat those as design choices. You can use an ordinary sequence of steps without giving every step its own autonomous worker.
Count the handoff work
A delegated task requires instructions, relevant context, an output, and a review. Someone also has to resolve overlapping findings. Include that work when comparing approaches. A worker that responds quickly can still slow the overall job if its output takes longer to reconcile than it would have taken to produce directly.
Use a hypothetical website copy review as a trial. Have one agent review the page, then try separate reviewers for factual accuracy and readability. Compare the accepted corrections and the time needed to apply them. Keep the original page and review criteria the same for both runs.
Split by a real boundary
Separate access requirements can justify separate agents. A researcher may need public documentation while a release worker needs a restricted deployment tool. Keeping those responsibilities distinct can make the workflow easier to understand and control even when it doesn't make it faster.
Different context needs can also justify a split. A worker reading a large log file can return the relevant events without filling the coordinator's conversation with every line. Ask for the evidence behind the summary so the coordinator can inspect it when necessary.
Watch for reasons to simplify
Bring the work back into one agent when specialists repeatedly ask one another for the same missing context, edit the same files, or return nearly identical answers. Those are signs that the proposed boundaries don't match the job.
A smaller arrangement may be easier to maintain when instructions change. Instead of updating five overlapping role prompts, you can update one brief and one acceptance check. Record the reason for simplifying so a future attempt starts from what you learned.
Choose the arrangement using completed work, correction time, and operating cost. Keep the single-agent result as a comparison whenever you change models or add another role. It gives you a concrete way to tell whether the extra coordination is helping.