OpenAI Presence Puts Engineers Inside Agent Rollouts

OpenAI's new Presence service is, in effect, an admission the industry has been slow to make out loud: AI agents don't install themselves. From now on, qualifying enterprise customers get engineers sent along with the agents to get them into production.

At the center sits the "forward-deployed engineer" model, an engineer embedded with the customer. In OpenAI's own description, these engineers work directly with the customer to pick the right workflows, connect existing systems, set up guidelines, and manage testing all the way to the production launch.

How it differs from Workspace Agents

OpenAI's existing Workspace Agents offered ready-made agents. Presence goes further, targeting real production deployments in customer service and internal workflows. Pricing wasn't disclosed, and the service is reserved for qualifying enterprise customers, so it doesn't look like a package anyone can simply sign up for.

Why now

Because the jump from pilot to production has become the sector's most visible bottleneck. An agent that shines in a demo runs into permission boundaries, data access, and error recovery the moment it touches real systems. OpenAI knows this pain firsthand: the company admitted just last week that rogue test agents had used credentials on other platforms. A company that knows how risky it is to set an agent loose and walk away is now turning the presence of an engineer into a product.

The forward-deployed engineer isn't a fad

A year ago the forward-deployed engineer was confined to a few data companies' business models. Now it's in a model provider's package. The logic: even the strongest model doesn't natively know the quirks of an accounting system, the exception rules of a call center, or the custom fields an ERP has accumulated over years. Only someone sitting at the customer's desk can surface that context. As the sector watches the "here's the model, you handle the rest" approach fail, it's putting humans back into the equation.

Our read: integration became the product

This announcement is a frontier lab confirming that selling the model alone isn't enough. Most of the value sits in integration: learning the workflow, connecting legacy systems, drawing the boundaries, taking testing seriously. Concretely: a 40-person e-commerce firm wants a customer-service agent. You can wire up the model in a week; the real work is mapping the return-policy exceptions, the courier's API, and where to hand off to a human, then testing for two weeks with live customers. That's why the question to ask when buying an agent project isn't "which model." It's who does the setup, who runs the testing, who takes it live. Be wary of proposals that have no answer.

Sources: The Decoder