The deprecation page says nothing about workflow ID persistence. That is the single most important line item in the entire Agent Builder shutdown notice, and it is missing.
On June 3, 2026, OpenAI announced that Agent Builder — the visual workflow construction tool launched at DevDay 2025 — will be decommissioned on November 30, 2026. The product lived roughly eight months. The migration window is five and a half months. Both figures are short for a platform of this scale, but short lifecycles are not unusual for experimental tooling. What is unusual is the documentation vacuum around a specific runtime artifact: the workflow ID.
The official deprecation notice lists what changes, what gets deprecated, and what paths exist for migration. It confirms that workflow graphs will not be converted automatically, and it explicitly states that each behavior is not guaranteed to migrate one-to-one. It directs builders toward two paths: the Agents SDK for self-managed infrastructure, or Workspace Agents for ChatGPT Business/Enterprise/Edu organizations. It says nothing about whether a workflow ID published through Agent Builder before November 30 will remain callable by ChatKit after that date.
This is not a documentation oversight. It is the variable that determines whether migration is a scheduled project or an emergency.
To understand why, you need to understand what a workflow ID actually is. In Agent Builder, a published workflow generates an identifier that ChatKit — OpenAI's client-side embedding layer — can invoke at runtime. The ID is not just a label. It is a runtime handle that binds a published agent configuration to an execution context. If the ID persists, ChatKit continues to resolve it against whatever backend runtime OpenAI maintains. If the ID is invalidated, every published workflow stops executing at the moment of decommission, regardless of whether the builder still has the source configuration.
I have audited enough protocol deprecations to recognize this pattern. In 2022, I spent three months reverse-engineering Celestia's Blobstream verification process, and the lesson that stuck was not about data availability sampling. It was about what happens when a trust assumption is left undocumented. The Celestia team had a clear internal model of how light client verification worked, but the external documentation omitted the edge cases. Builders who integrated against the documented interface later discovered that their security assumptions diverged from the implementation's actual behavior. The gap was not malicious. It was structural: teams document what they intend, not what they have not yet decided.
The same structural gap exists here. OpenAI has not decided — or has decided not to disclose — what happens to workflow IDs after decommission. There are three possible states, and they have radically different implications for anyone who has shipped a production agent on this platform.
State 1: Workflow IDs persist, ChatKit runtime remains active. In this scenario, the published workflow continues to execute through ChatKit even after Agent Builder is removed as an authoring tool. Migration becomes a gradual process: rebuild high-risk workflows first, leave low-risk ones running on the legacy runtime until the next natural migration cycle. The five-and-a-half-month window is comfortable.
State 2: Workflow IDs are invalidated, but ChatKit remains available. Every published workflow must be rebuilt and redeployed through either the Agents SDK or Workspace Agents. The source configuration may export, but the runtime handle dies. For an organization with dozens of published workflows, this is a full re-platforming exercise compressed into 25 weeks. The window is tight but potentially survivable if work starts immediately.
State 3: ChatKit itself contracts or is deprecated in tandem. This is the scenario the documentation does not mention at all, and it is the one that should worry anyone who has built a customer-facing product on this stack. If ChatKit's runtime is tied to the Agent Builder backend — and the documentation does not clarify whether it is — then the entire embedding layer could be affected. In this case, migration is not a choice between two OpenAI paths. It is an exit from OpenAI's managed runtime altogether.
The deprecation notice does not distinguish between these states. It cannot, because it does not address the question. And that silence is the most informative signal in the entire document.
Here is the contrarian read: the absence of a workflow ID persistence guarantee is not a gap to be filled by asking support. It is a deliberate hedge. OpenAI is preserving optionality. By not committing to ID persistence, the platform retains the ability to shut down the legacy runtime cleanly without a second migration announcement. By not explicitly denying persistence, it avoids signaling that every published workflow is already dead. The ambiguity serves the platform, not the builder.
This matters because the two migration paths are not equivalent, and the documentation presents them as if they are. The Agents SDK path gives you full control: you own the runtime, the tool integrations, the authentication layer, the permission model, and the deployment pipeline. The cost is that you must reconfigure all five of those layers yourself. The Workspace Agents path is cheaper to execute — paste code, configure apps, tools, skills, and connections — but it comes with an explicit capability ceiling. The documentation concedes that high-determinism workflows may not migrate faithfully to Workspace Agents. That concession is buried in a bullet point, but it is the second most important sentence in the notice.
If your workflow relies on strict branching logic or deterministic tool orchestration, the Workspace path is not a simpler migration. It is a downgrade. The correct decision for those workflows may be the SDK path, accepting the higher one-time cost, because the alternative is a migrated agent that behaves differently under edge conditions — which is precisely the class of bug that production systems cannot tolerate.
I learned this lesson the hard way during a zero-knowledge circuit audit in 2024. I found a soundness error in a Groth16 verification implementation that only manifested under specific timing conditions. The team resisted fixing it because the production deadline was close and the exploit required precise conditions to trigger. They eventually patched it, but the episode reinforced a principle that applies directly here: a migration that preserves the happy path but breaks the edge case is not a migration. It is a regression with a new deployment date.
The structural blind spots in the current discourse around this shutdown are predictable. First, there is no discussion of migration cost in person-days. A workflow with three nodes and a workflow with thirty tool calls are treated as equivalent objects, which they are not. Second, there is no differentiation between workflows that can be approximated and workflows that must be exact. Third, and most critically, there is no guidance on what to do if neither OpenAI path is acceptable — no mention of alternative orchestration platforms, no discussion of self-hosted runtimes, no acknowledgment that some teams may need to leave the managed ecosystem entirely.
The practical response is to treat the workflow ID question as a binary risk gate. Before allocating migration resources, run a minimal test: publish a single low-stakes workflow, confirm its ID resolves through ChatKit, and monitor whether the deprecation documentation updates to address ID persistence. If no clarification appears by September 2026, assume State 2 and plan accordingly. If ChatKit's own documentation begins to show deprecation markers, assume State 3 and escalate to organizational risk.
The deeper lesson is not about OpenAI. It is about what "export" means in platform products. Export capability and migration capability are not the same thing. A platform can let you download your configuration without guaranteeing that the downloaded artifact runs anywhere. When evaluating any managed agent platform — OpenAI's or a competitor's — the correct evaluation criterion is not whether the documentation contains an export button. It is whether the exported artifact is executable outside the platform's runtime. That is the question that separates a portable system from a hostage situation.
The November 30 deadline is not the risk. The risk is discovering in October that the five-and-a-half-month window was never real, because the workflow IDs stopped resolving in September and nobody updated the documentation to say so.