OpenAI Agent Builder is shutting down: timeline and migration options
OpenAI announced Agent Builder's deprecation on 2026-06-03. Evals become read-only on 2026-10-31, and Agent Builder, Evals and reusable prompts shut down on 2026-11-30. OpenAI's own path is the Agents SDK (code) or ChatGPT Workspace Agents (no code). If you want a visual workflow that runs on your machine instead, Mono Agent covers agent steps, approvals, branching, MCP and multi-agent runs, but has no hosted chat widget, no file search and no evals.
| Agent Builder | Mono Agent | Notes |
|---|---|---|
| Agent node | agent.ask | Runs on an agent CLI you already have (Claude Code, Codex, OpenCode, Copilot and others) with that CLI's own login |
| Guardrails / user approval | core.human_in_loop | Durable queue; the reviewer can edit fields before approving; a timeout auto-rejects |
| If/else | core.if, core.switch | Expression-based branching |
| MCP tools | Mono Agent MCP server; the agent runtime's own MCP servers | monoagentcli mcp exposes workflows to agents; agent.ask runtimes keep their configured MCP servers |
| Multi-agent | org.run, org.ask | Hand a task to an organization of agents |
| State / variables | core.set plus expressions | Values pass between steps as data |
| File search | No direct equivalent | There is no built-in vector store over uploaded files |
| ChatKit (hosted chat widget) | No equivalent | Workflows start from a webhook, schedule, the CLI or the desktop app |
| Evals | Not provided | Use the run history to review outputs; there is no grading tool |
The shutdown timeline
- 2026-06-03: deprecation announced.
- 2026-10-31: Evals become read-only.
- 2026-11-30: Agent Builder, Evals and reusable prompts shut down.
- Source: https://developers.openai.com/api/docs/deprecations/index.html (read 2026-10-13).
OpenAI's own migration path
OpenAI points Agent Builder users to two places. The Agents SDK is the code route: you keep OpenAI models and get full control over tools, handoffs and tracing, in exchange for writing and hosting the code. ChatGPT Workspace Agents are the no-code route inside ChatGPT. If your agent is a chat experience for people already in ChatGPT, or you want to stay on OpenAI's platform, one of these is likely the simplest move.
When Mono Agent is a better fit
- Your agent is really a workflow: a trigger, a few AI steps, an approval, an action.
- You want it to run on your own machine, with run history stored locally in SQLite.
- You want a person to approve or edit the output before anything is sent, with a queue that survives restarts.
- You would rather pay for AI through an agent CLI subscription you already have than through per-token API billing.
A starter workflow
Mono Agent ships a template at examples/agent-builder-migration.json. It is a starter, not a converter: a webhook trigger receives a request, agent.ask classifies it and drafts a reply, core.human_in_loop holds the draft for a person to edit and approve, and the approved reply is sent on. Replace the prompts with your Agent Builder instructions and add branches with core.if or core.switch where your agent had them.
There is no automatic import of Agent Builder exports. You rebuild each agent by hand, which takes minutes for a simple one and longer for agents with many branches.
What you lose
- Hosted chat UI: there is no ChatKit-style widget to embed on a website.
- File search: no managed vector store; you would call your own search service over HTTP.
- Evals: no datasets or graders.
- Hosting: Mono Agent runs on your machine, so that machine has to be on for schedules and webhooks to fire.
- Account: Mono Agent needs a monoes.me sign-in.
Frequently asked questions
When does Agent Builder stop working?
On 2026-11-30, together with Evals and reusable prompts. Evals become read-only earlier, on 2026-10-31 (https://developers.openai.com/api/docs/deprecations/index.html).
What does OpenAI recommend instead?
The Agents SDK for code-based agents, or ChatGPT Workspace Agents for no-code agents inside ChatGPT.
Can Mono Agent import my Agent Builder workflows?
No. Start from the examples/agent-builder-migration.json template and rebuild each agent; the concept table above shows which node replaces which.
Can I keep using OpenAI models in Mono Agent?
Yes, through an agent CLI that uses them, such as Codex, signed in with your own ChatGPT plan. agent.ask runs on whichever supported agent CLI you choose.
Is there a replacement for Agent Builder's evals?
Not in Mono Agent. It keeps a run history you can review, but it has no evaluation datasets or graders.