Key Takeaways

  • Modal launched its sandbox feature in 2023 as an experiment, but the product surged to a $300M run rate by mid-2025, accounting for one-third of the company's total revenue.
  • Sandboxes exploded because modern post-training and agentic loops require isolated, secure execution runtimes to evaluate code and score model performance.
  • Post-training thrives on short feedback loops such as unit tests and math proofs, whereas human tasks with multi-week horizons remain out of reach for reinforcement learning.
  • Software engineering tasks that complete within 24 hours to one week face rapid automation as models test and run code in parallel runtimes.

The Accidental $300M Product

In 2023, Modal launched sandboxes without fanfare. It was a side bet to let developers execute arbitrary code securely in the cloud. For nearly two years, the product moved slowly.

“We launched sandboxes in 2023. Didn't see a lot of success in the early days. I mean there was some success. It wasn't horrible, but it was a relatively insignificant part of our revenue up until summer of 2025,” (play from 34:41) Bernhardsson recalled. Then the curve bent vertical. “Sandboxes today is a 300 million run rate business. It's about a third of our revenue, and so that's been a crazy journey.” (play from 35:02)

What changed was not developer interest in running user scripts on SaaS sites. What changed was the AI training stack. When frontier labs moved beyond basic pre-training into aggressive post-training, they hit a hard infrastructure wall: models cannot learn to write reliable code simply by reading text tokens. They need to run the code, catch errors, check standard outputs, and update their behavior based on concrete feedback. Every step of that reinforcement loop requires an isolated sandbox that boots in milliseconds, executes untrusted code, and resets instantly.

The Limits of Reinforcement Learning

The rush into sandboxes reveals where AI works today, and where it stalls. Post-training is not magic dust you sprinkle over any business problem. It demands deterministic verification.

“I think post-training works because there's a lot of tasks that people do in real life that have fairly finite feedback mechanisms. Coding is a great example,” (play from 42:45) Bernhardsson noted. A compiler returns an exit code. A test suite passes or fails. A math proof resolves or breaks. Because the feedback loop is immediate and binary, models running in sandboxes can generate millions of attempts and learn from each failure.

That dynamic creates a clear divide for builders. If an agent can run an experiment, inspect the result, and iterate in minutes, AI will master that workflow fast. “A lot of people ask me what's going to happen with software engineers in the long term,” (play from 43:42) Bernhardsson said. “And I think we're probably going to automate everything that an agent can do within 24 hours or a week.” (play from 43:46)

Tasks with month-long feedback cycles, such as hiring an executive, setting product roadmaps, or running a brand campaign, do not offer clean verifiers. Reinforcement learning cannot optimize a loop it has to wait forty days to measure. The sudden demand for sandboxes proves that post-training works best where the digital world gives instant, automated answers.

What to Do With This

Audit your product backlog by feedback duration. Separate the engineering tasks that take less than 24 hours from projects requiring cross-team consensus over weeks. For every short-horizon task, set up automated tests and an isolated test environment this week so an agent can write, run, and verify pull requests without human oversight.