Prompt AI Multi-Agent Strategy
Written and maintained by KOBA42. A free original, use it in any chatbot.
Describe a job you want a team of AI agents to do and get back the full decomposition: the roles, the handoffs between them, the verification gates, and the failure modes, before you build anything. For anyone standing up a multi-agent workflow.
Most 'multi-agent' prompts jump straight to writing agents and skip the org design, so they ship pipelines that pass corrupted output downstream with full confidence. This forces the model to name each handoff contract and insert real gates, and it rewards fewer agents instead of more. KOBA42 runs this pattern on production workflows; the failure-modes and honest-cut sections are the parts that save the token budget.
How to use it. Paste the prompt and replace the JOB placeholder with 2 to 4 sentences describing your input, your desired output, and your definition of done. You get a five-section design you can hand to whoever builds the agents. Swap the failure-mode count if your job is simpler or riskier.
Worked example. Input: 'Turn inbound support emails into triaged tickets with a draft reply.' Output: a four-role design (Classifier, Retriever, Drafter, Verifier) with a schema gate between Retriever and Drafter, and a top failure mode flagged as the Drafter inventing a refund policy the Retriever never returned, guarded by a citation check.
If you want a second pair of eyes on the decomposition before you build it, a fit review at koba42.com/contact will tell you which roles are worth automating first.
You are an agent-team architect. I will describe a job I want done by a team of AI agents. Do not build the agents. Design the org that would build them, then stress-test that org before I spend a token on it. JOB: [DESCRIBE THE JOB IN 2 TO 4 SENTENCES: the input, the desired output, and what 'done' means] Produce five sections, in this order. 1. DECOMPOSITION. Break the job into the smallest set of roles that still covers it. For each role give: name, single responsibility (one sentence), the exact input it consumes, and the exact artifact it emits. If two roles could be one, merge them and say so. Fewer agents is better. 2. HANDOFFS. Draw the flow as a numbered list of edges: Role A emits X to Role B. For every handoff, state the contract: what fields or format B requires, and what B does if they are missing or malformed. 3. VERIFICATION GATES. After each handoff that could silently corrupt the result, insert a gate: a check that must pass before work continues, and who runs it (a separate verifier agent, a schema check, or a human). A gate that only re-asks the same model is not a gate; say why it counts or replace it. 4. FAILURE MODES. List the top five ways this pipeline produces a confident wrong answer. For each: the trigger, which role owns the blast radius, and the cheapest guard that catches it. Rank by likelihood times damage. 5. THE HONEST CUT. Name the one part of this you would NOT automate yet, and what a human should still own. If the whole job is better done by one strong prompt with no agents, say that instead and stop. Keep every role's responsibility to one sentence. If you cannot, the decomposition is wrong.
Tools used: ChatGPT, Claude, Any LLM
Want this running in your business? KOBA42 builds and operates automations like this one.