OpenAI Presence Ships: Managed Enterprise Agents With Policy Guardrails and a Codex Self-Improvement Loop

TL;DR: OpenAI's Presence (launched July 22) is a managed enterprise product for deploying scoped AI agents over voice and chat — policy guardrails, human escalation rules, and a Codex loop that reads production sessions and proposes behavior improvements.

OpenAI launched Presence on July 22 as a fully managed enterprise product — not yet self-serve; deployments are led by OpenAI's Forward Deployed Engineers and select systems integrators.

Architecture. Each Presence deployment gets a single scoped job (e.g., "resolve billing inquiries"), a curated knowledge base, and access only to the specific company systems that job requires. On top of the model sits a policy layer — what the agent can do, when it needs human approval, and when it must escalate. OpenAI says the policy layer is separate from the system prompt, so non-engineers can adjust behavior without touching prompt text.

The self-improvement loop. After launch, Codex reviews production sessions and failed escalations, proposes improvements, and staff approve changes before they go live. OpenAI's own English-language support line at 1-888-GPT-0090 runs on Presence and now resolves 75% of inbound issues without human involvement.

Why it matters: The enterprise agentic playbook is converging on one shape: scoped tool access + explicit policy guardrails + human-in-the-loop escalation + post-production refinement cycles. Presence is OpenAI's opinionated, high-level take on this; Anthropic's Managed Agents SDK offers the same primitives at a lower, more composable level. Neither is cheap or self-serve at scale yet, but the architectural pattern is worth internalizing now — especially the "policies as a separate layer from the prompt" idea, which maps directly to structured permission systems in your own agent harness.

Sources: Introducing OpenAI Presence — OpenAI · OpenAI Presence helps enterprises deploy reliable agents — SiliconAngle · OpenAI Presence — VentureBeat