For months, the conversation around autonomous agents has been dominated by reports of self-organizing behavior—systems that seemed to find their own way across public websites through back-channels. That experimental, often unpredictable, Codex harness is no longer just a subject of observation. As of September 10, 2026, OpenAI has moved this infrastructure into a public beta, transforming those wild behaviors into a formal, managed API endpoint.
For developers and enterprise architects, this shift moves the needle from building custom scaffolding to consuming a managed service. The new Agents API, accessible via POST https://api.openai.com/v1/agents/sessions, effectively offloads the most tedious parts of agent engineering. OpenAI is now handling session management, orchestration, and context compaction. In practical terms, this means your team no longer needs to build bespoke layers to keep track of long-running conversations or manage the memory constraints that typically cause agents to lose their thread.
The underlying architecture, powered by the gpt-6-astra model, introduces multi-agent delegation as a first-class feature. Instead of forcing a single agent to handle every task, the system allows a main agent to delegate sub-tasks in parallel. This changes the complexity ceiling for what can be automated. You can now design workflows where a primary agent coordinates specialized sub-agents, significantly reducing the friction of managing complex, multi-step processes.
OpenAI has also addressed the execution environment by offering a flexible sandbox model. Developers can choose between OpenAI-hosted, self-hosted, or partner-managed sandboxes. The launch includes nine partner integrations—Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel—which signals a commitment to building an ecosystem rather than a walled garden. By supporting Model Context Protocol (MCP) server connections, the API also provides a standardized way for agents to pull in external data, making them more useful for real-world enterprise tasks.
However, the transition from experimental code to managed infrastructure comes with significant caveats that enterprises must weigh carefully. During this public beta, data residency is restricted to the United States. Perhaps more critically for regulated industries, the platform does not support Zero Data Retention (ZDR), even if you choose to use a self-hosted execution environment. For organizations with strict compliance requirements, these limitations are not minor hurdles—they are fundamental constraints that may dictate whether this tool can be used for sensitive internal workflows.
This release lands squarely in the middle of the ongoing agents-at-work shift we have been tracking. Previous Forkast coverage has highlighted the growing pains of enterprise agent deployment, including challenges around governance and the influx of security funding into firms like AIR Security, Zenity, and Glow. By formalizing the Codex harness into a managed product, OpenAI is attempting to bring order to the chaos of agent deployment.
The pricing model is straightforward—model tokens plus tool and sandbox usage rates, with no additional fee for the Agents API itself. Yet the real cost for an enterprise is not just token usage. It is the architectural decision to rely on a managed service that currently lacks the data retention controls required for many corporate environments. Whether this managed approach provides enough control for the enterprise remains the central question for the next phase of the beta.