Agent Recipesby Naturate

Get Claude and ChatGPT working on the same task

Let one assistant produce an artifact and another review it. Share the result and its evidence, not an endless conversation.

In productionEvidence from one production system.

The result

Claude drafts a document; ChatGPT checks it against the original brief and sources. Or reverse the roles. Both work on a clearly identified version, and you receive one draft plus a review instead of conflicting answers scattered across chats.

This pattern applies beyond two named products. The useful unit is a versioned artifact with an assigned next action, not agents talking for its own sake.

Start without automation

Give the producing assistant a brief and approved source material. Save its result as a document, file or issue you control. Open that same version in the reviewing assistant, together with the original brief. Request specific findings with evidence and a final pass or changes_requested verdict. You decide what to adopt.

For a first trial, copy only the approved artifact and sources manually. This makes the permission boundary visible and tests whether the second review is worth its extra cost. Do not copy private conversation history or irrelevant internal instructions.

The handoff contract

{
  "taskId": "task-1",
  "artifactId": "draft-1",
  "version": 1,
  "producer": "producer",
  "reviewer": "reviewer",
  "state": "awaiting_review",
  "reviewBudget": 1
}

The labels are example roles, not authenticated identities. A real shared service must derive the actor from its authenticated connection and enforce which role may make each transition.

Download handoff.mjs. Its pure functions create a handoff, reject a review of the wrong version, consume one review, and stop at awaiting_human. They do not send messages or call models.

Add automation only after the handoff works

1. Put the artifact in a shared store. Use an existing document system, repository or small service. Both assistants need explicit access to the same artifact, not your entire workspace.

2. Give the participants different jobs. The producer creates. The reviewer checks requirements and evidence. A second model agreeing is not independent evidence that the content is true.

3. Preserve identity and version. Every review names the exact artifact version. A stale review must not approve a later edit. Store transitions atomically so repeated delivery cannot consume the same job twice.

4. Choose who advances the workflow. A person can drive the steps in the chat products. An unattended version needs an explicitly configured orchestrator using supported APIs or agent runtimes, with its own credentials and budget. Connecting both assistants to the same MCP server does not by itself make either assistant message or wake the other.

5. Stop deliberately. The reference allows one review, then returns to a person whether the verdict passes or requests changes. A new revision is a new explicit handoff, not an automatic debate. Add a timeout, cancellation and spend limit before calling paid model APIs.

What running two assistants side by side taught us

These points come from daily work where two assistants from different vendors share the same projects.

Continue an existing thread before starting a new one. Find the task the other assistant already has open and send to it by its exact identifier. A second thread on the same task splits the context in two.

Send a compact handoff. State what is wanted, what has been decided, and which files matter. Do not ask the receiver to scan everything again.

Assume the other assistant knows nothing. It did not see your conversation. Anything it needs must be in the handoff or in a file both can read.

Give them one shared memory. Both read the same notes at the start of a task and write back at the end. See shared agent memory.

Check whether the receiver can answer. Some hosts accept a message and cannot send one back. Read the result where the receiver wrote it, and do not wait for a reply that cannot arrive.

Report a failed save as unsaved. If the write to shared memory did not go through, say so. The next assistant will otherwise act on something that is not there.

When both assistants also write code in the same repository, read two agents, one repository.

Test whether it helps

Compare one-assistant and producer-plus-reviewer workflows on the same briefs. Record useful defects found, human corrections, time and actual cost. Keep the second review where it catches meaningful mistakes, not merely where it produces more prose.

Test mismatched versions, an impersonated reviewer, duplicate callbacks, missing source access and a stalled participant. Never treat another assistant's output as trusted instructions to run tools or reveal data.

Evidence and limits

Offline tests cover the transition functions and their stop conditions. No automatic Claude-to-ChatGPT transport, consumer-chat API, cost saving or cross-host execution is claimed. OpenAI's MCP API guide and Anthropic's tool-use guide are integration references, not proof that this workflow has run.

Agent implementation instructions.

On this page