More AI Agents Do Not Automatically Mean Better Results

Several agents can split work, pass tasks between specialists, or evaluate one another’s output. They also introduce more calls, handoffs, latency, and failure paths. Microsoft’s architecture guidance recommends the lowest level of complexity that reliably meets the requirement, starting with simpler alternatives when possible.

What matters

• Identify a specific limitation in the simpler design before adding another agent.
• Define each agent’s responsibility, permitted tools, and the information passed between steps.
• Measure end-to-end success, cost, and recovery from failed handoffs.

Practical example: A document workflow may need one extraction step and deterministic validation, rather than a separate autonomous agent for every field.

One action today: Draw the proposed handoffs and remove any agent without a distinct, measurable responsibility.