OpenAI shipped Presence on July 22 — a managed platform for running production agents, starting with realtime voice and chat. It bundles the unglamorous parts of taking an agent live: policies and standard operating procedures, connections to your systems, guardrails, approved actions, a simulation tool that manufactures rare edge cases, evaluations, and a Codex-powered loop for improving the thing over time.

If you only read the feature list, you'll file this as "OpenAI ships an enterprise agent console" and move on. Don't. The feature list is table stakes — Microsoft, Google, and a dozen startups ship versions of each piece. The signal is buried one line down in the availability note.

The one-line read: OpenAI just priced the operations layer around an agent — policy, guardrails, evals, improvement — as the product, and it's selling it white-glove. You can't buy Presence with a credit card.

1. What Presence actually is#

Presence is not a library and not a visual builder. It's a managed service: OpenAI (and select systems integrators) stand up the agent, wire in the guardrails, and run the improvement loop with you. The most interesting included piece is the simulation tool — it generates the rare, adversarial requests an agent will only hit occasionally in the wild and measures how the agent responds, so you find the failure before a customer does. Guardrails don't just block risky outputs; they constrain how the agent is allowed to touch the systems it already has permission to reach. That's the part naïve integrations always skip.

This is the same production discipline we've watched the whole field converge on — the quarter the coding agents spent shipping guardrails instead of horsepower was the same tell. Presence is that instinct, productized and sold.

2. The tell: white-glove, not self-serve#

Here's the sentence that matters. Presence is limited general availability, delivered through OpenAI's Forward Deployed Engineers and partners — not self-service. There is no console you swipe a card into. OpenAI hasn't published a price, a geography, or the cost of the integration work.

A company that built its entire business on self-serve API access — swipe a card, get a token, ship — just launched its agent platform as a consulting-shaped, human-delivered engagement. That's not a rollout detail. That's a statement that the hard part of agents isn't the model call; it's the policy, the connections, and the eval loop, and that OpenAI thinks those are worth sending an engineer to install.

3. What it signals for founders#

Two things, and they point in opposite directions — both useful.

The value moved up the stack. For two years the model was the product; everything around it was your problem. Presence inverts that. The base model is now a component; the sellable thing is the ring of operations around it. If you're deciding where to build, build in the ring — guardrails, simulation, evals, observability and trace review — not another thin wrapper on a completion call. OpenAI just told you where it thinks the money is, and it's the same place the enterprise agent spend is flowing.

White-glove delivery is a confession: OpenAI thinks the operations layer is too hard to self-serve. For the 99% of builders who'll never get a Forward Deployed Engineer, that's not a wall — it's a spec.

The self-serve lane is wide open. Presence is enterprise-only and human-delivered by design. Every solopreneur, every seed-stage team, every mid-market shop that will never rate an FDE still needs exactly this stack — and now has a reference architecture for it, courtesy of OpenAI's own feature list. If you sell any slice of agent-ops, your market didn't shrink; it got a lighthouse. (We wrote the build-it-yourself version so you don't have to reverse-engineer the announcement.)

4. The one date on your calendar#

Presence changes nothing you're running today unless you're an enterprise engaging OpenAI directly. But it arrived alongside a deadline that does: OpenAI's AgentKit Agent Builder and the Evals product retire on November 30, 2026. If either is in your pipeline, that's a real migration, and we've mapped where to move. The pattern is consistent — OpenAI is trading the assemble-it-yourself tools for the run-it-for-you service. Read the trade, then decide which side of it you're building on.