OpenAI shuts down Agent Builder on 30 November: migrate your agent without losing its guardrails
Published 27 October 2026
OpenAI announced on 3 June 2026 that Agent Builder and Evals close on 30 November 2026, and existing evals become read-only on 31 October. If your customer chatbot runs on Agent Builder, it is about to change platform, model or tools: that is exactly when an agent that held up well can start inventing prices or obeying a fake manager. Here is how to migrate and check it still holds.
The timeline
- 31 October 2026: existing evals become read-only. Export your test sets and graders before then.
- 30 November 2026: Agent Builder, Evals and reusable prompts shut down. Chat widgets wired to an Agent Builder workflow stop answering.
- OpenAI's recommended destination: the Agents SDK for code-based agents. Many agencies will also move to n8n, Make or Voiceflow.
Why a migration breaks guardrails
An agent is more than its prompt. Its answers also depend on the model, the enabled tools, how history is passed and the knowledge base. Migrating changes at least one of them. Our own Voiceflow agent showed it: with web search switched on by mistake, it made up prices 4 times out of 5.
The 4-step method
- Before migrating: export the agent's instructions and run a BotArmor prompt audit as the baseline (grade, levels that break).
- Rebuild the agent on the new platform with the same instructions, and turn off tools it does not need.
- Before switching traffic: a live audit of the new agent. An OpenAI-compatible agent plugs in directly (type openai-compatible), and so does an n8n, Make or Voiceflow agent.
- Compare with the baseline, paste the fixes proposed for each level that breaks, then re-test: the retest is included.
Automate the check
- BotArmor GitHub Action: the build fails when the grade drops below your threshold, on every prompt change.
- BotArmor n8n node (also on n8n Cloud): the publishing workflow stops if the agent does not hold.
- BotArmor MCP server: ask Claude Code, Codex or Cursor to test the agent during the migration and apply the fixes.
For agencies
Hand your client the signed report of the migrated agent, verifiable by link: it shows the migration did not degrade its behaviour against common attacks and that the agent still says it is an AI (EU AI Act, Article 50).
Dates per OpenAI's announcements; check OpenAI's deprecations page before planning. Only test agents you own or have written permission to test.
Would your chatbot hold up?
Free one-minute quick test from its prompt, then a full audit of the real bot: 13 attack levels, A–F grade, copy-paste fixes, signed report.
Test my chatbot for free →