Agent Builder is the authoring and governance surface within AgentFlow. It gives authors a structured way to create, test, and publish AI agents — and gives approvers a governed workflow to review and control what reaches end users.

How Agent Builder works

An agent author writes a system prompt, selects the tools the agent is allowed to use, and tests the agent interactively before submitting it for review. When a publish request is submitted, the system automatically scans the system prompt for security violations. If a violation is detected, the request is auto-blocked before it reaches a reviewer. Valid requests move to a reviewer for approval. Once approved, the agent is published and immediately available in Discover Agents. End users can find it in the catalog and start a conversation with it right away.

While a request is pending review, the agent is locked. If the author needs to make changes, they can withdraw the pending request at any time. Withdrawing returns the agent to Draft state and unlocks the editor. If an agent already has a published version, that version remains live while the new version is under review.

System prompt security scanning

Every publish request triggers an automated scan of the agent's system prompt. The scan evaluates the prompt for security violations. If a violation is detected, the request is auto-blocked before it reaches a reviewer.

The same scan runs when an author tests a draft agent. This is a build-time check — runtime conversation guardrails do not re-check the system prompt on each conversation turn.

For more information, see System prompt security scan.

When to use it

Use Agent Builder when you need to:

• Create and iterate on a custom AI agent
• Test an agent before submitting it for review
• Apply a governed review process before an agent reaches users

For more information about creating and publishing an agent, see Create, test, and submit an agent for review.