Introducing OpenAI Presence

Source: OpenAI Date: 2026-07-22 URL: https://openai.com/index/introducing-openai-presence

Summary

(Direct fetch of the source page returned 403; this summary is reconstructed from secondary coverage and search results, not the primary page.) OpenAI launches Presence, an enterprise product for deploying scoped AI agents that answer questions, use company systems, take approved actions, and escalate to a human when needed. Each deployment is job-specific (billing resolution, insurance claims, IT service requests) with access limited to what that job requires, and the company sets policy for what the agent can do autonomously versus what needs approval or handoff. It ships as a white-glove offering — OpenAI Forward Deployed Engineers and select systems integrators lead deployments; it’s not yet self-serve.

Implications

Agent-runtime-hardening thread: Presence’s design — narrow per-deployment scope, explicit policy/guardrail layer, mandatory escalation path — is the enterprise packaging of the same “constrain the room the agent runs in” principle that’s been showing up in Claude Code’s trust/permission model and in mise’s inert-reader mode. That the FDE-led, non-self-serve rollout is the chosen go-to-market (rather than an open API) suggests OpenAI doesn’t yet trust the guardrails enough to hand them to arbitrary customers unsupervised — worth watching whether self-serve access loosens those constraints or the constraints hold and only the delivery model changes.