OpenAI Agents API makes agent orchestration a managed service
The public beta can remove a large chunk of production-agent plumbing, but it also concentrates state, cost and portability risk in OpenAI’s control layer.
What launched
On September 10, 2026, OpenAI released Agents API in public beta. It manages sessions, orchestration, context compaction and recovery, while agents can run in an OpenAI-hosted sandbox, on a team’s own infrastructure or through a supported partner.
The task economics
There is no separate Agents API fee: teams pay the selected model, tool and hosted-sandbox rates. The build decision should therefore turn on total cost per completed task—including subagent fan-out, retries and human intervention—not the headline token price.
What we would do
Virtual Arc would move one bounded internal workflow first and retain our own interface around tasks and tools. During public beta, we would not make it the control plane for a customer-facing system or use it for workloads requiring non-US residency or Zero Data Retention.
OpenAI has not merely added another agent SDK; it has made a bid to own the operating layer between the model and production. That is useful because sessions, context compaction, recovery, sandboxing and subagent coordination are exactly where internal agent prototypes become expensive to maintain. It is also where lock-in becomes hardest to unwind. Virtual Arc would test Agents API now, but only behind our own task contract and tool boundary. We would run one bounded workflow—incident investigation, repository triage or document review—against our existing harness, measuring completed-task cost, wall-clock latency, intervention rate and reproducibility. We would not migrate a customer-facing control plane during public beta, and we would exclude workloads requiring non-US residency or Zero Data Retention, which the service does not currently support. If it cuts operational toil without degrading those numbers, we would buy the harness; if not, we would keep the orchestration portable and wait.