Chinese
Uncategorized

Agent Builder Is Shutting Down: Move the Workflow Before the Date

OpenAI has set a shutdown date for Agent Builder. The practical job now is moving workflows, state, and acceptance checks into code you can review.

Jacky Wang 4 min read 78 views

I ran into this kind of problem while tidying up an agent workflow: the model was not the scary part. The scary part was a useful canvas with no obvious exit. OpenAI’s current Agent Builder documentation says the product is being deprecated and is scheduled to shut down on November 30, 2026. Existing users can keep using it during the transition. ChatKit remains available, and OpenAI recommends the Agents SDK for workflows that need to continue as code.

That is not an argument against visual builders. A canvas is a good way to get the first path working: collect input, call a tool, ask for approval, and return a result. Production systems need more than a visible path, though. They need typed boundaries, retry semantics, permission ownership, and a clear answer to the question “what happens if this runs twice?” If the only representation is a diagram, migration becomes archaeology.

Turn every node into an explicit contract

My first pass would not be a rewrite. I would make an inventory. For each node, record its input schema, tools it may call, retryable failures, and whether it can cause a side effect. A screenshot is not enough. It cannot tell a reviewer whether a value is an enum or a free-form string, or whether a retry can create a duplicate invoice.

Even a small text record helps:

node: fetch_order
input: { order_id: string }
side_effect: none
retry: timeout, 429
approval: no

The Agents API guidance is useful here because it treats an agent as a combination of instructions, tools, state, and orchestration. That decomposition exposes rules that were previously hidden in the builder. I would turn each important node into a named function or agent, with the contract kept beside the implementation.

Keep the interface away from execution

If the workflow is embedded in a web product, ChatKit can remain the interaction layer while the execution logic moves into the SDK or a service you control. That separation matters. Changing a chat surface should not quietly change authorization.

For a Kubernetes deployment, I would also give the worker its own request ID, timeout, structured event log, and idempotency key. A browser request should not stay open while an agent waits on a slow external system. The worker should be able to resume or fail in a visible state.

I would keep a small fixture set and compare the old and new paths: a normal request, missing fields, an upstream 429, a changed tool response, and an approval refusal. Compare tool calls and side effects, not only the final prose. A nicer answer that sends an email twice is still a failed migration.

For each fixture, I would save the input, the expected tool boundary, the approval decision, and the final state. The comparison can then be run in CI without calling a real customer system. For tools that cannot be safely replayed, use a stub that records the requested operation and verify that the request is made once with the expected idempotency key. This also makes a rollback meaningful: the team can restore the previous worker and know which in-flight jobs are safe to resume.

Treat the date as a test deadline

The real work before November 30 is not dragging the same boxes somewhere else. It is proving that the system can deploy, roll back, and be audited without the original canvas. Put the workflow definition in git, secrets in the runtime secret store, and acceptance fixtures in tests. Before switching, export a non-production run record and confirm every tool call maps to a new function or agent. Check that logs do not retain sensitive prompt input, keep the old version as a read-only comparison path, and give the new executor explicit timeouts and stop conditions. Run the new path in shadow mode if the side effects are expensive.

Visual orchestration will not disappear. It will become more useful as a prototype and a communication surface. The production boundary still wants reviewable files, reproducible inputs, and explicit authority. If I had a workflow in Agent Builder today, I would migrate the node with the largest side effect first, not wait for the final week.

I would also ask a second person to review the migration table. Missing a hidden approval or an implicit state transition is easier to catch in a short contract review than after the canvas is gone.

Sources: Agent Builder documentation, Introducing the Agents API, OpenAI Agents SDK documentation

发表评论