Cafiyn Pulse
← Startup Signals
5 min readUpdated 2026-09-17Issue 3

OpenAI's Agents API will hold your agent state. What breaks if you ever need to leave.

Two major providers moved into the agent state layer in the same week. A managed harness is a real shortcut, as long as you keep a copy of what it holds.

ByKarthik KumarCafiyn Innovations
At a glance
Launched
Public beta, 10 September 2026
Managed by OpenAI
Session orchestration, compaction and recovery
Data residency
US only, per the docs
Zero Data Retention
Not supported
Previous managed layer
Assistants API, shut down 26 August 2026

What launched

On 10 September OpenAI released the Agents API in public beta. OpenAI describes it as a managed harness: it runs session orchestration, context compaction and recovery, sessions persist across turns, and agents can run in OpenAI-hosted or self-hosted sandboxes. Model usage is billed at the chosen model's rates, with tools and hosted containers at their standard rates.

Two constraints in the documentation matter for anyone selling to regulated or non-US customers. Data residency is currently available only in the US, and Zero Data Retention is not supported.

The part that already happened once

The Agents API launched fifteen days after OpenAI shut down the Assistants API on 26 August. The Assistants API was also a layer that held agent state on the provider's side: threads, runs and the tool lifecycle around them. Teams that built on it had to migrate that state to the Responses and Conversations APIs before the shutdown date.

That is not an argument against the new API. It is a description of what vendor-held state costs when the vendor changes direction. The migration was manageable for teams that had kept their own record of conversations, and painful for teams whose only copy lived in the provider's threads.

This is an industry direction, not one product

In the same week, on 14 September, Anthropic added on-demand compaction to its Messages API in beta. You send a compaction parameter and the API returns a signed compaction block that summarises the messages you sent, which you then send back in place of those messages on later requests.

Both features are useful. Both also produce agent state that only one provider can interpret. A signed summary block from one vendor cannot be handed to another vendor's model. A session held in one provider's harness cannot be resumed on a different provider. The trend is clear enough to plan around.

Keeping a portable copy

The practical answer is to write your own vendor-neutral record alongside whatever the provider holds, as the session happens rather than after the fact. It does not need to be elaborate. It needs to be complete enough that you could reconstruct the conversation for a different model if you had to.

  • Store every message and tool result in your own database as it streams, keyed by your own session id
  • Record which model, which tools, and which system prompt produced each turn
  • Keep agent definitions, tool schemas and prompts in your repository, not only in a provider dashboard
  • Treat provider-signed summaries as a cache that can be rebuilt, not as the source of truth
  • Test a resume on a second provider once, so you know what actually breaks rather than guessing

What to do

  1. 1If you adopt the Agents API, write every session item to your own store as it streams.
  2. 2Do not promise customers Zero Data Retention or non-US data residency on top of it until the documentation changes.
  3. 3Keep agent definitions and prompts in version control.
  4. 4Run one test that resumes a real conversation on a different provider, and record what fails.

Common questions

What is the OpenAI Agents API?

A managed harness for agents, released in public beta on 10 September 2026. OpenAI runs session orchestration, context compaction and recovery, and sessions persist across turns. Agents can run in OpenAI-hosted or self-hosted sandboxes.

Does the Agents API support Zero Data Retention?

Not currently. OpenAI's documentation says it does not support Zero Data Retention and offers data residency only in the US.

How is it different from the Assistants API?

Both hold agent state on OpenAI's side. The Assistants API shut down on 26 August 2026, with the Responses and Conversations APIs named as its replacements. The Agents API is a newer managed harness launched fifteen days later.

What did Anthropic launch the same week?

On 14 September 2026 Anthropic added on-demand compaction to the Messages API in beta, under the compact-2026-09-04 header. It returns a signed compaction block that you send in place of the summarised messages on later requests.

Sources

Read against the primary documentation rather than secondary coverage. Where a figure comes from a provider's own docs, it is quoted as published on 2026-09-17.

Related

What the Assistants API shutdown broke →Model what a provider switch would cost →OpenAI's full shutdown calendar →

More from Signals

1.8 million Android apps were scanned for hardcoded keys. Here is how to check yours.
Finding secrets in shipped apps is now a cheap, automated pipeline. Anything compiled into an APK, an IPA or a JavaScript bundle should be treated as already published.
Meta's Muse agent can now check out at your store. What merchants should change.
Agent purchases are arriving through wallets and single-use cards, not through merchants adopting a new protocol. Many merchants will receive them without choosing to.
What Brevo's SSO incident teaches about multi-tenant SAML and hijacked sending accounts.
One incident, two lessons: why email authentication did not help, and the tenant-scoping mistake anyone shipping SSO can make.
Cloudflare now sorts bots into Search, Training and Agent. What it means for your site and your agent.
The same change lands on two audiences. Site owners can now accidentally block search; agent builders will find more doors closed.
OpenAI has published seven shutdown dates. Here is the calendar.
Seven dates between 23 October and 26 February. Every one announced months ahead, in public. The failure mode is never the notice.
Nvidia bought Hugging Face. Your open-weight fallback is now a vendor relationship.
The commitments say nothing breaks. The structural change is that "we can always self-host" is now a dependency on one vendor rather than on a neutral commons.
The Videos API goes away on 24 September and OpenAI names nothing to move to.
Every other deprecation on OpenAI's page offers a migration target. This one does not, which makes it a different kind of problem.
Whisper shuts down on 26 February. The risk is not the deadline.
Five months is a generous window. Whisper is also typically wired into a background job that fails into a queue rather than into a user's face.
The Assistants API is gone. Your app may not have told you yet.
It shut down on 26 August 2026. The replacement is not a drop-in, and the most common migration mistake changes your model's behaviour without throwing an error.
One Copilot code review costs 13 requests. Here is the rest of the maths.
Copilot now bills on usage rather than seats. The headline change is simple. The multipliers underneath it are where teams get surprised.
Row-level security is on by default now. That is not the same as being protected.
Lovable fixed the default. The gap that actually leaks data was never the default, it was the policy, and no platform can write that for you.
Your model comparison is measuring the wrong number.
Every price list is per token. Your bill is per finished piece of work. Those two numbers can rank models in opposite orders.
An agent with production log access is a new identity. Treat it like one.
Always-on agents that read production logs are arriving fast. The access-control question they raise is old and well understood, which is the good news.

Open the tool.

Ten production categories, three minutes, and a ranked list of what to fix first.

Check your own stack