OpenAI Presence: Why Production Agents Still Need Engineers

OpenAI already runs its own English-language phone support line on Presence, its new enterprise agent platform — and the system resolves 75% of inbound calls without a human touching the case. The catch: you can’t sign up for it yourself.

What Happened

OpenAI introduced Presence in late July 2026 as a deployment platform for running AI agents in production — specifically customer service and internal service workflows across voice and chat, not the chat-based “GPTs” of Workspace Agents, which OpenAI shipped earlier this year for internal team automation. Presence is a different layer: it bundles policies and standard operating procedures, guardrails, approved actions, pre-launch simulations, and evaluation tooling into a single pipeline for getting an agent from prototype to live traffic, as The Decoder reported.

Two things separate Presence from a typical agent-building toolkit. First, the testing layer runs simulated conversations against common requests, edge cases, and higher-risk scenarios before anything goes live, scoring outcome correctness, policy adherence, tool use, and escalation decisions. Second, a Codex-powered review process keeps running after launch — it reads production sessions and escalations, then proposes behavior changes, though a human still has to approve every change before it ships, per OpenAI’s own announcement.

None of this is self-service. Presence rolls out through a limited general-availability program, and access runs through OpenAI’s account team, OpenAI’s Forward Deployed Engineers, and select systems integration partners who help select workflows, connect existing systems, and manage the process through to production launch.

Why It Matters

The gap Presence is built to close is a familiar one on this beat: agents that work in a demo and fall apart against real traffic. It’s the same failure mode behind Stack Archive’s coverage of why AI agent projects stall before reaching production, and behind the observability vendors building tooling specifically to catch agents behaving unpredictably once they’re live.

What’s notable about Presence is that OpenAI isn’t selling a fix for that gap as a feature toggle. It’s selling the fix as a service relationship — simulation before launch, guardrails that intervene when an agent drifts outside its lane, and a standing improvement loop after go-live, delivered with engineers attached rather than handed over as a self-serve dashboard. The 75% resolution figure is OpenAI’s own dogfooding result, not an independent audit, but it’s a concrete number tied to a real, ongoing production workload rather than a benchmark run.

What This Means for Builders

The FDE-and-partner delivery model is the real signal here. It tells you OpenAI doesn’t think “give teams an SDK and guardrail docs” is sufficient to get agents reliably into production yet — the hard part is still the integration, policy definition, and testing discipline around the model, not the model itself. Compliance details are also thin: OpenAI has mentioned “trust mechanisms” without specifying how Presence handles requirements like the EU AI Act, which matters if you’re evaluating this for a regulated workflow rather than a general support line.

For teams outside OpenAI’s qualifying-customer tier, the practical takeaway isn’t “wait for Presence” — it’s that the simulate-guardrail-monitor-improve loop Presence formalizes is buildable with existing agent frameworks and observability tooling today. Presence is a preview of what a fully packaged version of that discipline looks like once a frontier lab productizes it.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.