OpenAI opens managed agent harness to developers
Reviewed by Helen Jones
OpenAI announced the Agents API in public beta on 10 September 2026. It exposes a managed cloud harness based on the infrastructure behind Codex, including tool use, persistent environments, context handling, and subagent coordination for long-running work.
Until now, a team wanting an AI agent for recurring work had to assemble the operational layer around the model themselves: files, tools, context, retries, and a way to keep work moving after one chat ends. That is where one-off prompting tends to break when a task needs more than a single answer. The new API does not turn a vague process into a safe workflow. It makes the harness more available, so the real management question becomes clearer: which work may the agent touch, what information and tools does it need, and where must a person check or decide before the work continues.
Analysis
Pick one recurring task that currently starts as a blank AI chat. Write its inputs, tool or data access, expected output, and the one point where a human must approve or correct it before comparing agent options.
Source note
Pulse published by Collab365 Spaces, reviewed by Helen Jones on . Cite as "OpenAI opens its managed agent harness to developers", Collab365 Spaces.