OpenAI's Agents API Exposes the Codex Harness as a Managed Service

OpenAI has opened up the runtime behind Codex as a hosted product called the Agents API. Instead of building their own orchestration loop, developers hand OpenAI responsibility for session state, context compaction, task recovery, and delegation to subagents, while the application supplies the tools and picks where code actually runs. Agents operate inside a sandbox — OpenAI-hosted or self-hosted — where they can execute commands, edit files, load skills, reach external systems through MCP servers, and emit artifacts. The API is organized around four primitives: the agent (model, instructions, tools, MCP servers), the environment, a durable session, and the stream of events and items that flow in and out.

The pitch is aimed squarely at long-running, tool-heavy automation: the published examples include an incident-response agent that investigates alerts and gates recovery actions behind human approval, a read-only SQL data analyst, and a GitHub bug reproducer. Multi-agent mode lets a session fan work out to concurrent subagents, and sessions can be steered mid-turn or resumed later without rebuilding conversation context. Billing stacks model API rates, tool rates, and container rates for hosted sandboxes.

The security and compliance caveats are the part worth flagging. The Agents API is US-only for data residency and does not support Zero Data Retention — and OpenAI explicitly notes that choosing a self-hosted sandbox does not make it ZDR-eligible. Teams running agents with live code execution and MCP-connected access to internal data should weigh that retention posture, plus the usual sandbox-escape and tool-permission concerns, before wiring one of these into production infrastructure.