From copilot chat to governed agents

On September 10, OpenAI released its Agents API (opens in a new tab) in public beta. It exposes the managed harness behind Codex: durable sessions, orchestration, context compaction, tool and MCP connections, and a choice of where code actually runs, whether that is an OpenAI sandbox, your own infrastructure, or a partner.

Google made a similar structural move earlier in the year when it folded Vertex AI into the Gemini Enterprise Agent Platform (opens in a new tab), pitched as one place to build, scale, govern and optimize agents.

The direction is consistent. We are moving from chat copilots to hosted agent runtimes with governance attached.

What this genuinely fixes

The plumbing under a long running agent is real engineering, and most teams were building it badly.

Keeping session state across hours or days. Compacting context before it overflows. Recovering when a step fails halfway. Coordinating subagents. Deciding what the agent is allowed to touch. If you have written any of that yourself, you know it is a lot of unglamorous code that has nothing to do with your business.

OpenAI is charging nothing extra for the harness itself, only for the tokens, tools and containers underneath. That is a meaningful shift in where the work sits. Before, the harness was yours to maintain. Now it can be somebody else's.

What it does not fix

None of it tells you which process to point the agent at.

That is not a criticism of the release. It is a statement about which problems are technical and which are organizational. A hosted runtime removes the reason not to start. It does not tell you what the approval rule is when an invoice arrives with the wrong purchase order number, or who is allowed to override it, or what an auditor will want to see next year.

We keep meeting teams who read an announcement like this and conclude they were blocked on infrastructure. Usually they were blocked on a description of their own work.

Governance is not only a platform setting

Both platforms lead with governance, and both mean something specific and narrow by it: identity, policy, access control, guardrails on what an agent can call. That is necessary.

It is also worth reading the fine print, because beta means beta. Coverage of the Agents API notes that during the public beta the service supports data residency in the United States only and is not eligible for Zero Data Retention, and that choosing a self hosted sandbox does not change that, because OpenAI still holds the session layer. If you work in a regulated setting, that is a procurement conversation, not a footnote.

The wider point is that platform governance and process governance are different things. A platform can enforce that an agent may only read from one system. It cannot tell you that a human has to sign off before a claim gets denied. That rule lives in your business, and somebody has to write it down.

The starting move has not changed

Our advice to a mid-market team without an in-house AI group is the same as it was before any of these launches.

1. Pick one process that someone internal owns and is willing to sit with
2. Map the exceptions before the happy path
3. Ship into a real week rather than a sandbox
4. Keep the consequential decision with a person
5. Measure the result against the hours it was supposed to remove, review time included

Notice that none of those five steps are affected by which runtime you chose. Pick the platform your team can actually operate, then spend your attention on steps one, two and four, because that is where projects succeed or quietly fail.

Better tools raise the cost of skipping the thinking

There is a version of this news that reads as pressure to move faster. We would read it the other way.

When the runtime was hard, a badly chosen project died in development and wasted a few weeks. When the runtime is easy, a badly chosen project ships, touches real records, and takes longer to unwind than it took to build. Lower friction raises the cost of not knowing what you are automating.

The teams that will get the most out of a hosted harness are the ones who already did the boring work: they can name the hours, they know where the exceptions go, and one person owns the outcome. That work is not automated by any of this. It is just made more valuable.

Tell us where the hours go

No pitch required. Name the team and what the work costs them in a week. That is enough to tell whether we are useful.

Contact