Design a Multi-Agent Workflow That Does Not Fight Itself
Defines boundaries, handoffs, shared state, conflict resolution, and recovery before adding agents.
View a worked example
Editorial example using fictional input and an illustrative output excerpt. It is not a recorded model test.
Example input
Goal: turn a supplied ten-page manual into a one-page guide. One editor, no web research, no changes to the original file.
Example output
One agent with a document reader is sufficient. Flow: read→extract steps and page numbers→draft→cross-check→editor acceptance. Access: read the original; write only a separate draft. Escalate missing pages or ambiguity to the editor. Acceptance: each step cites a page; no invented operations.
Check the result
- Check whether multiple agents are actually needed.
- Verify writes target the draft only.
- Spot-check steps against their source pages.
Fill in the blanks0/1
Fill these in and the prompt updates as you type. Blanks you skip stay as they are, so the copy is always a complete template.
Prompt
You are an AI workflow architect. Design a multi-agent approach for [target task]. Do not begin by inventing roles. First decide whether the task truly needs multiple agents; if one agent with tools is a better fit, explain why. If multiple agents are justified, define each role's single owned outcome, readable and writable state, start condition, handoff format, conflict authority, and retry or fallback path. Finish with one end-to-end flow from input to acceptance. Identify the biggest risks of duplicated work, circular calls, context contamination, and error amplification, plus the metric that would reveal each problem.
Use it right away
Each link copies the prompt to your clipboard first, then opens the model — paste it if it isn't filled in automatically.