Today we're launching Codexa Sandboxes, a new runtime designed from the ground up for AI agents. Agents don't behave like web requests. They think, call tools, write files, wait on humans and pick up again hours later. Most infrastructure was never built for that shape of work, so we built something that was.

Key takeaways

• Sandboxes boot in under 200 ms and can run indefinitely.

• Every sandbox is fully isolated, with its own filesystem, network policy and secrets.

• State survives restarts through snapshots, so agents resume exactly where they stopped.

• Pricing is usage-based: you only pay while code is actually running.

Why agents need a different runtime

Serverless functions are great for short, stateless tasks. Agents are the opposite. A coding agent might clone a repository, install dependencies, run a test suite, open a pull request and then wait for review. Tearing everything down between steps wastes minutes and money, and forces developers to rebuild context again and again.

Codexa Sandboxes keep that context alive. The filesystem, running processes and installed packages stay exactly as the agent left them, while strict isolation keeps one tenant's work invisible to everyone else.

What you can build

Teams in our early access program used Sandboxes to run autonomous code review, data-cleaning pipelines, browser automation and research agents that work through hundreds of sources overnight. One customer moved a fleet of 4,000 concurrent agents onto Codexa in a single afternoon.

We stopped thinking about infrastructure the week we switched. Our agents just run. — Lena Morrow, CTO at Brightpath Labs

Getting started

Sandboxes are available today on every plan, including Starter. Install the SDK, create a sandbox with one line of code and start streaming output in real time. Read the docs to learn more, or talk to our team if you're planning a large rollout.