OpenAI's July 22 launch of Presence makes a useful shift visible: deploying an agent is becoming a managed operating problem.
Presence combines policies, standard operating procedures, guardrails, approved actions, simulations, evaluations, escalation rules, system connections, and a post-launch improvement process. This is OpenAI's description of its own product, not independent proof of business outcomes. The architecture still matters because it shows where the enterprise burden is moving.
Production starts where the demo ends
A demo proves that an agent can complete a task under selected conditions. Production has to answer harder questions.
What information can the agent use? Which systems can it reach? Which actions can it take without approval? When should it stop? Who receives the escalation? What happens when the policy, product, customer behavior, or connected system changes?
OpenAI says each Presence deployment begins with one specific job and only the knowledge and access required for that job. The company sets the policies, approval points, and human handoff rules. Simulations and graders then test outcomes, policy compliance, tool use, and escalation behavior before release.
OpenAI also reports that Presence resolves 75% of inbound issues in its own English-language phone support channel and that its improvement loop reduced human handoffs by 15 percentage points in ten days. Those are vendor-reported results from OpenAI's own environment. OpenAI does not publish enough methodology in the announcement to treat them as a forecast for another business.
The improvement loop needs a human owner
The most important part of the announcement appears after launch.
OpenAI says production sessions, escalations, and quality signals expose gaps. Codex can investigate those signals and propose changes. Teams test each proposal against the production version and approve a controlled rollout.
That loop does not remove human responsibility. It makes responsibility more specific. Someone still has to decide which signals matter, which evaluation represents acceptable performance, which change is safe, and when the agent should be narrowed, paused, or rolled back.
An agent without that operating rhythm can stay technically online while drifting away from the work it was designed to handle.
Set the operating model before choosing the product
OpenAI Presence is available through a limited general-availability program for eligible enterprise customers. Deployments are led by OpenAI Forward Deployed Engineers and selected systems integrators; it is not a self-service product.
Most companies will evaluate other products as well. The operating questions remain useful regardless of vendor:
- Define one job, its inputs, its permitted actions, and its boundaries.
- Name the owner for quality, policy, access, and escalation decisions.
- Establish evaluations for normal work, edge cases, and higher-risk conditions.
- Make the human handoff visible and test whether it receives enough context.
- Review production signals on a fixed cadence instead of waiting for complaints.
- Keep a controlled path for testing, approval, rollout, and rollback.
The launch is one event. Agent operations are the system that keeps the work reliable after reality changes.
Related services: Implementation, Optimization, Support