Skip navigation
OpenAI's new Agents API is a hosted Codex runtime that can connect to sandboxes like E2B.
0:00 / 3:08

OpenAI's new Agents API is a hosted Codex runtime that can connect to sandboxes like E2B.

Vasek Mlejnsky

@mlejva8.9K followers

Visit on X
91 likesDownloadOriginal post

23,210 views

OpenAI's new Agents API is a hosted Codex runtime that can connect to sandboxes like E2B. OpenAI manages the model, harness, and agent sessions, while code executes on an isolated E2B machine with its own filesystem and preinstalled tools. We built an agent workbench where each chat runs on its own machine. It starts from a custom template, auto-pauses between turns, forks to try another approach from the same starting point, and can even act on videos inside the sandbox. Thanks Ondrej for the workbench demo! Try it: e2b sbx create openai-agents-api-python-sdk Build on top of the workbench: Docs:

10 Comments

Comments from the original post on X (Twitter)

  • Daniel Liao's profile picture

    @Daniel_Bernath_

    @OpenAI @e2b LETS GO! Big Day for all the devs out there

    1 like

  • 安叫兽|Bird🕊️ 🔶 BNB's profile picture

    @ajs6888

    @OpenAI @e2b 这组合起来终于像能干活的 agent 了

    1 like

  • John Rood's profile picture

    @johnroodepic

    @OpenAI @e2b auto-pause + fork is the important part. the durable primitive isn't a long-running machine. it's replayable task state that can hydrate into a clean machine, branch, and resume.

    1 like

  • Vext Labs's profile picture

    @VextLabs

    @OpenAI @e2b hosted harness, isolated execution — the right split

  • Michael Waitze's profile picture

    @MichaelWaitze

    @OpenAI @e2b The way it handles orchestration and long-running sessions is actually quite impressive.

  • catman's profile picture

    @catmanyau

    @OpenAI @e2b auto-pausing each sandbox between turns could make long agent sessions much cheaper; how quickly does a resumed machine come back?

  • gleb's profile picture

    @Peaceful_HN

    @OpenAI @e2b The isolated filesystem plus preinstalled tools is the part that matters most in practice. Most of my agent failures used to come from environment drift, not the model. Good to see that getting solved at the infra layer instead of every project rebuilding it.

  • Lawrence Tang's profile picture

    @LawrenceTangAI

    @OpenAI @e2b Auto-pause between turns is the product, not a missing automation. A hosted harness that keeps going with no human gate is just a runaway script with better branding. Forks only help if the failed branch can’t write back by default.

  • Ben Mo's profile picture

    @benmodev

    @OpenAI @e2b The fork is the part I'd like to see explored more. A side-by-side view of what each attempt changed would make "try another approach" much easier to judge than scrolling two chat histories.

  • Shlok Madhekar's profile picture

    @shlokbuilds

    @OpenAI @e2b how's your cold start time on spinning up a fresh e2b sandbox per session?