OpenAI's Agents API hands you the Codex harness

OpenAI now runs the agent loop for you: sessions, sandboxes, subagents, and context compaction behind one API.

13 cards · 3m · tap to begin

From · · 3m

OpenAI's Agents API hands you the Codex harness

OpenAI now runs the agent loop for you: sessions, sandboxes, subagents, and context compaction behind one API.

In brief

OpenAI now runs the agent loop for you: sessions, sandboxes, subagents, and context compaction behind one API. You bring the tools and pick the execution environment; OpenAI runs everything else, but only in the US and never with zero data retention. Originally reported by OpenAI Developers.

Codex, now as an API

The Agents API opens up the same Codex harness through an OpenAI-managed endpoint. You call it, OpenAI runs the loop.

“The Agents API gives your application access to the Codex harness through an OpenAI-managed API.”

Who owns what

OpenAI handles sessions, orchestration, context compaction, and recovery. Your app supplies the tools and picks where the agent actually runs.

A sandbox with real hands

Agents can execute code, edit files, talk to MCP servers, and leave behind artifacts. This is a working machine, not a chat box.

“Agents can operate in a sandbox where they can execute code, edit files, connect to MCP servers, and produce artifacts.”

Three separate meters

Model usage bills at the chosen model's API rates, OpenAI tools at their standard rates, and OpenAI-hosted sandboxes at container rates.

Start from a working example

Run a directory-tree script in a hosted sandbox, or have subagents compare release notes and merge their findings into one answer.

What people build with it

Reference apps cover incident response with approval gates, a Slack investigator, a read-only SQL data analyst, a GitHub bug reproducer, and a policy-driven document reviewer.

Four concepts, that's it

Agent is the model plus instructions, tools, and MCP servers. Environment is the sandbox. Session is a durable running instance. Events and items are the input and output.

The shape of a session

Create a session and OpenAI provisions the environment. Send a task, stream progress or wait on a webhook, then send more work or steer mid-turn.

What the harness does for you

Running commands, applying skills, reaching external data via MCP, accepting steering mid-task, summarizing old work, delegating to subagents, and resuming later.

Capabilities are config

One create call wires it all up: a model, instructions, a tools array with programmatic tool calling, an MCP server, and web search, plus multi_agent settings.

“"multi_agent": { "enabled": true, "max_concurrent_subagents": 4 }”

Bring your own box

The environment block can point at a self-hosted sandbox with your own workspace directory and capability directories for skills.

State survives the turn

Sessions retain state, so you continue work without rebuilding conversation context. Delete sessions and published artifacts when you no longer want them around.

“The Agents API retains session state so you can continue work across turns without rebuilding the conversation context.”

Read the data rules first

US data residency only, and no Zero Data Retention. Running your own sandbox does not change that, so regulated workloads need a closer look before you build.

“Choosing a self-hosted sandbox does not make the Agents API ZDR-eligible.”

What to remember

You bring the tools and pick the execution environment; OpenAI runs everything else, but only in the US and never with zero data retention.

Read the original on OpenAI Developers

React

Sign in to react and comment.

Comments (0)

Life is short. Keep it sweet. Respect others' opinions and be kind!

    Recommended next

    More decks on openai and related topics.