OpenAI’s Agents API Brings the Codex Harness to Developers in Public Beta
A new API exposes versioned agent infrastructure — context compaction, tool search, parallel tool use — so apps inherit harness upgrades with each model launch.
1 min readCTRL Staff
On 10 September 2026, OpenAI introduced the Agents API in public beta: the same style of harness that powers Codex, offered as a flexible API for third-party agents.
What you get beyond “call the model”
OpenAI’s post highlights context management that compacts earlier turns as sessions approach limits, tool search that loads relevant tool definitions on demand, and support for parallel tool calls and multi-agent style decomposition. The harness is described as open-source Codex-derived logic that OpenAI operates while developers can inspect the public codebase.
The pitch is durable: when models improve, you should not rewrite your entire agent loop. Versioned harness access is meant to keep application code stable while behaviour improves underneath.
Product reality check
Agent APIs succeed or fail on reliability, cost, and debuggability. Compaction and tool search are answers to known failure modes — context bloat and tool-definition sprawl — that every serious agent builder has hit.
CTRL’s take: this is infrastructure competition as much as model competition. Whoever makes long-running, tool-using agents boring to operate will win more production traffic than whoever wins a single leaderboard screenshot.
Comments
Loading comments…