AIBRIEF
Codex is OpenAI's coding agent brand; the harness is the runtime that manages sessions, tools, and recovery. The new API hands that runtime to outside developers as a managed service.
OpenAI's new Agents API hands the runtime behind its internal coding agents to outside developers. OpenAI's developer documentation describes it as a managed harness: OpenAI runs the agent's sessions, orchestration, context compaction, and recovery, while the developer's application supplies the tools, the MCP servers, and the choice of execution environment.
The split packages long-running agent work (keeping a session alive, compacting context, recovering from failed tool calls) as an API. OpenAI-hosted sandboxes can execute code, edit files, and connect to MCP servers. Pricing follows the underlying model's API rates plus standard container rates for those sandboxes.
OpenAI's docs ship example applications to ground the abstract: an incident response agent, a Slack bot, a read-only SQL data analyst, a GitHub issue investigator, and a document reviewer. Those examples point to a target buyer who wants a known-good agent runtime and is willing to run inside OpenAI's container to get it.
A Hacker News thread on the launch flagged a tradeoff. Commenters argued OpenAI's harness is locked to its own proprietary frontier models, while non-lab competitors can pick the best-fit model per workload.
The docs do not cover deeper pricing tiers, rate limits, or enterprise security scopes. Those gaps are worth a follow-up read before any procurement call.