Go from idea to a working agent faster with the Agents API. Build and run cloud agents with the Codex harness, fully managed by OpenAI. We handle orchestration, long-running sessions, and context management. You focus on what makes your agent unique. Available in public beta.

OpenAI runs the agent loop on its infrastructure, coordinating model calls, tool use, and context. You control the agent’s capabilities and choose where it runs code and works with files.

Your sandbox, your call. Bring your own sandbox or connect a sandbox provider. Choose the environment that fits your workload: • CPU, GPU, and memory options • Fully managed environments or deployments within your VPC • File and secret storage options We offer first-class integrations with @blaxelAI, @CloudflareDev, @daytonaio, @digitalocean, @e2b, @modal, @OracleCloud, @RunloopAI, and @vercel.

We’re also introducing OpenAI-hosted sandboxes. Your agent can: • Run code • Work with files • Produce artifacts Supply files, install packages, and add skills and plugins while we provision and manage the sandbox.

there are local sandboxes too, you know?

Wow, this is amazing. I wonder who worked on this...

Now bring back the thing people are actually asking for, just in case you haven’t noticed people are asking for 4o it’s the best model for chatting. #4oForAll

Can we use our ChatGPT subscriptions as inference? :)

is this like Claude Managed Agents?

cool, but I want 4o as a choice too. hopefully if the word choice doesn't upset you then give users the choice for 4o as well. many people value that model. #4oForAll

Managed orchestration is the part I'd actually pay for. I would still want to see it hold a long session on a messy real task, not the launch demo.

Want to save a ton of money? Use and put @opencode agent cli in there, running on free or cheap models from @OpenRouter, e.g. DeepSeek V4.1 Flash.

@e2b @atduarte

How long is long-running?

Managing long-running sessions natively solves half the state corruption pain. The real test is how well it handles cross-agent dependency graphs without snowballing context costs.

Amazing

This is a real game changer. But what about cost??? Versus just making your own agent using Astra?

Sandbox deployed. Adult supervision sold separately. 💀

Goblin mode

Now in harness business, Model business is restricted by compute, but this wont be a case with this niche

Managed orchestration and long-running context is the boring half of agents; the product bet is what uniqueness you can ship on a stable harness.

🤩

long-running sessions are the claim. does the managed Codex harness keep tool credentials alive across hours, or do you re-auth each turn?

every new "agents api" is just another abstraction layer on top of the same llm that u now gotta pay extra for. the product doesnt make the model better it just monetizes the complexity the company itself created around "agents" in the first place

Agents API + Codex harness in public beta. they handle the long sessions.

No extra fee for the API itself, you pay for tokens and tools. Subagent count you cap yourself with max_concurrent_subagents. Compaction is the automatic one, it fires as a session nears the context limit and picks what survives. On a multi hour run that is what sets the bill. Any way to control when it triggers?

Managed orchestration is the real pitch, not the harness itself.

I've been burning through my API credits from agentic 3D work may I get some credits to try some longer running 3D modeling experiments in the sandbox until I'm able to arrive at a more stable harness?

Managed orchestration + long-running session state matter more than model choice here - most agent failures come from context handling, not raw intelligence. #AgenticAI #ContextEngineering @OpenAI

Agents are great, but why not put an entire AI employee behind an OpenAI-compatible endpoint?

Agent API?

Managed orchestration and long-running Codex sessions are a big unlock. The next layer is multiplayer: multiple humans and agents working in shared channels and threads, with durable inboxes and human approval for consequential actions. That’s what we’re building at @puffoai.

This looks good…but I guess there must be plethora of open source alternatives…

long-running sessions without explicit state management hooks usually hide weird failure modes

@grok summary ?

This is killer! We need a min sandbox integration asap! can't wait to play around with this more.

Orchestration is the right thing to take off people's plates. Running my own agents daily, almost none of my failures are the model. They're handoffs, permissions, and a job that quietly never started. Managed sessions help only if I can see into them.

Managed sessions are the right trade until one dies mid-run. If the harness dies between two tool calls, does it resume from the checkpoint or replay the call? That answer decides whether a payment or email step can live in there at all. What does the beta guarantee?