Chew's second build on this page, and the first one to put numbers on snapshot-and-fork. The agent is deliberately small: read a buggy Python file, ask DeepSeek to fix it, run assertions on what comes back. The measurement is the point. A cold sandbox plus dependencies plus agent code costs 11 to 14 seconds per worker, so a 12-worker fleet pays roughly 147 seconds before any of them does useful work. Snapshotting the parent once takes 15.8 seconds, each fork after that lands in about 2.4 seconds, and the same fleet comes up in around 45 seconds. He then deletes files inside one fork to confirm the other eleven never see it.
Read it for the caveats as well. He hit a concurrency ceiling lower than the free tier advertised and a file write to a root-level path that failed, which we have since fixed in the docs. His own conclusion is the honest one: forking earns its keep for batch evaluation and RL rollouts, and not for a single long-lived agent.
We didn’t write this one — it’s Chew Loong Nian’s piece, published on Towards AI. The note above is ours; the full article is theirs.