Sample content for editorial review. Examples are illustrative.
Start with the work, not the diagram
A diagram with several agents can look like a plan before the task itself is understood. I would first describe the workflow as a sequence of inputs, decisions, and outputs. Ask which steps require model judgment and which can be ordinary code. A deterministic validator does not become more useful because it has been given an agent name and a conversational interface.
Try a bounded single-agent workflow
Give one agent a narrow goal, a limited set of tools, and clear stopping conditions. Measure whether it can complete representative tasks with reviewable evidence. Record failure points, tool calls, latency, and cost. That baseline matters because a more elaborate design should be compared with something concrete. Complexity is a tradeoff to evaluate, not a sign that a system has become more capable.
Add a second role for a reason
Separate agents may help when the workflow needs distinct contexts, responsibilities, or permission boundaries. A drafting role and an independent checking role can be worth exploring, but independence needs more than a different role description. Define what evidence the checker receives, what it can challenge, and how a disagreement is resolved. Otherwise the second role may simply repeat the first response more confidently.
Treat handoffs as interfaces
Specify the information passed between roles and the conditions that make a handoff valid. Use structured outputs where they help validate required fields. Keep an owner for the overall task so a conversation between components does not continue without a useful result. Put limits on repeated handoffs and provide a clear route for the system to stop or ask for human input.
Keep the simpler option available
Compare the designs on the same evaluation examples, including failure recovery and review effort. A multi-agent workflow may be useful for one part of the product and unnecessary elsewhere. If the additional role does not improve a measured outcome, remove it. The engineering question is whether separate responsibilities produce better evidence and more dependable behavior than the simpler system you already understand.
