Offload: Speed up the agent loop by running tests remotely

bfogelman

we’ve been using it internally (on sculptor) and the speed ups are crazy — we can now have our agents run tests all the time and iterate quickly! Excited other people can now give it a spin

nvader

I'm one of the devs of Offload here, happy to answer any questions. We built Offload because parallel coding agents exposed a bottleneck we hadn't fully felt before: agents write fast now, but our integration suite (345 Playwright tests) were still taking 12 minutes per run. There’s only so much parallelism you could get out of a single laptop, and our agents were spending more and more time running and waiting for tests. Offload is a Rust CLI, and agents invoke it directly to spin up isolated Modal sandboxes and distribute tests to them. The results are merged back and returned to your agent. The integration test suite for Sculptor makes heavy use of Playwright, and we saw the time for a run drop from ~726s to ~120 (More than a 6x speedup!) We've deliberately chosen a pluggable architecture so that we can operate with almost any test framework. We ship with support for pytest, vitest and cargo-nextest out of the box. We take a similar approach to support any bash-scriptable sandbox provider, but we’re launching with support for Modal to start. There are some caveats to share: suites under 2 minutes won't see meaningful gains since the overhead of preparing the testing image and sharing out the sandboxes is going to dominate. Short-running unit tests at high volume also get less scheduling efficiency from LPT. We're looking at work-stealing as a next step there. We'd love to hear any feedback! Install with `cargo install offload`, and the code is at github.com/imbue-ai/offload