Behind every great agent is a great stack. We’re getting the best one together for a night that’s stacked AF. Join @e2b, @FireworksAI_HQ, and @braintrust for technical talks, drinks, and bites on 9/30. RSVP:

What we need is a smart orchestration. Matthew's take is correct. Agents don't always need a full sandbox (which should be a VM, not a container btw), but it's not the agent's job to figure that out. It's the job of the underlying infrastructure. Frankly, given where we're headed, we shouldn't even be thinking in terms of one sandbox per agent. Agents will need their own private clouds. This is one of the problems we're thinking about. If working on this interests you, you should reach out.

Back in 2023 we were telling everyone that every agent will need a computer. We built a whole company based on this thesis. Wow look where we are. The fastest growing consumer agents have their own computers.

I found one of the older photos of me and @co_valenta working together. Roughly 2012 at my grandpa's who had a spare space for us, we were building a game if I remember correctly

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_sh for the workbench demo! Try it: e2b sbx create openai-agents-api-python-sdk Build on top of the workbench: Docs:

Give @Meta's Muse Code an @e2b sandbox and its native video understanding yields an efficient parallel video-processing pipeline. We handed the same demo recording to Muse, Claude Code, and Codex, each in its own E2B sandbox, and asked for a step-by-step breakdown. Muse Spark 1.3 takes the video natively. The others have to shell out to ffmpeg, extract frames, and re-send every image on every turn. In this clip, Muse finished up to 9.3× faster at up to 99.4% lower cost. Spin up one isolated sandbox per clip, and Muse applies edits across hundreds of videos in parallel. Try it: e2b sbx create muse Docs:

Most companies integrating AI, including many new startups I see here on X, are building around what models can do today. Which makes them be behind on the spot But once you build the thing for now, it's already 3 months old. By the time you build for today’s capabilities, your product is already three months out of date. By then, there’s a new model or emergent behavior enabling use cases that will become the main thing another three months later. So the thing you decided to build is effectively 6 months behind before you even ship it. To build a company that survives this pace of change, you need to at least: A. Continuously revise your company’s vision. B. Ensure your product enables behaviors that will become prevalent six to twelve months from now.

@GavinSBaker I think it’s likely that an enterprise harness will be short lived . Agents should be able to assemble a task specific harness as sandboxing tech evolves by following @E2B approach including BYOC ..

Paper Instruments benchmarked sandbox providers for their RL rollouts, and E2B's tool execution came in up to 3x faster than every other sandbox provider they tested, reducing idle GPU time and training cost. @paperinstr trains frontier models for knowledge work (consulting, finance, banking, law), running thousands of concurrent GRPO rollouts, each needing the solver and grader fully isolated from one another to prevent reward hacking. Every rollout also needs to start from the exact same state in order to reflect the policy being trained accurately. It also needs to boot and execute fast, since the GPU sits idle waiting on the rollout to finish. With @e2b, Paper Instruments isolates each rollout in a microVM, builds the initial state from a template, then snapshots the running sandbox, letting thousands of rollouts launch straight from the snapshot instead of booting cold, so they get consistency and speed while reducing training cost. Read the full case study:

Cursor x E2B Many of our customers use E2B in their internal development workflows. Some work in regulated industries where code and internal services can’t leave the company. For teams that need control over the execution environment, Cursor’s cloud agents can now use E2B as their machine. With Self-Hosted Machines, @cursor_ai handles the agent loop while @e2b runs the code. Teams can give agents bigger machines and start them from custom E2B templates, so their Cursor workloads boot straight into an environment with internal dependencies already in place. Try it:

No matter what usage or revenue chart I look at it's up and to the right again and again. We’ve grown our revenue dozens of times over in less than a year. Extraordinary times yes but also an effect of compounding work of the @e2b team.