The Agents API Makes the Harness the Product

On September 10, OpenAI put the Agents API into public beta. The pitch is quiet and important: the thing that made Codex useful at scale was not only the model. It was the harness — context compaction, tool search, sandboxes, and subagents that can run for days.

Most teams I talk to still glue this together themselves. A new model drops, the prompt breaks, the tool list overflows the window, and a week disappears into runtime work. A versioned harness is OpenAI admitting that the runtime is now part of the model release.

Three ideas worth stealing even if you never call the API

1. Tool search instead of a giant toolbox

Loading every function definition into the prompt is how agents get expensive and confused. Search for the next tool the way you search a codebase. Bring in only what the current step needs.

2. Programmatic tool calls

Let the agent write a short program that fans out calls, filters results, and returns a digest. The model should not re-read 400 rows of JSON just to pick two.

3. Subagents with a clock

Long jobs need isolation. A research subagent, a coding subagent, a reviewer. Each gets a workspace and a time budget. The parent only sees the artifact.

const run = await agents.start({
  goal: "Summarize last week's support tickets by theme",
  tools: ["db.query", "docs.search"],
  subagents: [{ role: "cluster" }, { role: "write" }],
  compact: "on_overflow",
});

The open-source Codex harness is the other half of the story. You can read how they coordinate tools even if you host your own model. That is the educational win: agent quality is becoming an engineering discipline, not a prompt file.

If your agent dies when the model card changes, you built a demo. If the harness survives the next launch, you built software.