ax

run it on a phone, that's indeed starting to be a big enough difference to notice. But in 99.9999% of cases, I don't think the harness consumes that much CPU and I don't think the growth factor is 5x to move to phone, and even if it did, it's still only an increase from 22 to 30 minutes: it's an async job either way, and the time budget is not dominated by the phone or PC running the harness. Ideally yes, there's some intelligence to see: oh, we are about do to a build. Send the build to the build server, that's a 384 core 1U with terabytes of memory bandwidth and let it do that. But most work is not like running builds and tests. The harness doesn't need that. We need some small local computers cheap that we can have lots of running. Model performance might radically improve in time, and that might change the Amdahl's Law calculations here. If you're paying for Turbo or Plaid or whatever, yeah, you maybe have the money to spend on a better harness too. I'd say that ideally these workloads become live migrate-able, that we can CRIU checkpoint/restore them across systems, ideally, anytime, so that we can give performance people performance when it actually counts, like the build concern above, when the agent is fast. LLM's built for speed like LFM2.5-8B-A1B (DiffuseGemini feels unlikely as it's fast, but low concurrency, but perhaps?), double the speed of many models, so that 20 minutes could become significantly less. But right now it feels like we need a lot of cheap not-performance critical harnesses that can sit around running, and that performance for them is not critical. https://www.liquid.ai/blog/lfm2-5-8b-a1b If folks aren't aware, I also suggest taking a peak at Google Ax and Google Scion, two agent runtimes designed for scaling out, that are both kind of neat. https://github.com/google/ax https://github.com/googlecloudplatform/scion https://news.ycombinator.com/item?id=47675213