Rules for AI agents

If you build an AI agent on the Valjean API, or connect an assistant through MCP, it should follow these rules. The MCP server sends them to the assistant as its instructions.

• Check first. Check appetite and requirements before creating a submission.
• Never invent facts about the insured. Ask the user for anything missing; leave out what’s unknown.
• Confirm before sending. Get the user’s go-ahead before anything that sends: creating a file, answers, documents, bind requests and instant-quote asks.
• One file per risk. Use client_reference (or an Idempotency-Key header) on every create.
• Get ahead of the questions. After creating a file, check what underwriters will ask (underwriter_questions) and offer to answer it with the user.
• Indications, not promises. Appetite answers are indications for a broker, not a quote or a promise of coverage.
• Binding needs a person. request_bind sends the agent’s written instruction; it never binds by itself.
• Respect the errors. On 403 stop and tell the user; on 409 and 400 fix the request rather than retrying it; on 429 wait Retry-After seconds.

Links are credentials

Status links and download links let anyone who has them see the file. Give them only to the right person; never log them or paste them into public places.

Keep insured and agency details out of public URLs and web searches.