4 mins critical path, 1h30m wall clock time (bunch of optional tests)
was around 40 mins on the critical path at the start of the year. gains came mostly from sharding tests, replacing webpack with vite, eslint/prettier with oxlint/oxfmt, caching node modules
My experience has made me more interested in seperating deterministic checks the judegment based reviews.Authentication changes have also taught me that passing inidividual checks doesnt mean an integration works end to end.
Trigger a review agent for any commit which has an open PR and has passed build. Use open weights, or at minimum a different model family from the one which wrote the code. I will be publishing a write up soon(tm), inspired by but different from, cloudflare's post.
I would napkin math the averaged cost per-review to be less than $1, which is less than our most rigorous validation pathway in CI (needs GPU for up to an hour). Time-wise around ~10-15m, with variability - and a bug in Argo with pod scheduling delays. This is fine because developers check in on their own PRs when they do and then make the changes that are needed. A key point here is that we are keeping humans in the loop in the PR.
This is a large monorepo, I would expect much lower numbers for polyrepo setups.
Fun to see this topic, I just wrote about how I got mine to 4 minutes, optimization details in my blog: https://alexitc.com/blog/2026-10-23-from-27-minutes-to-4-cut...
More or less, I did:
- Avoid unnecessary work
- Don't boot more often than needed
- Parallelism in tests
- Caching rules
My post has an ongoing live survey with ~570 responses:
- 30% says 5-15m
- 26% says >30m
- 23% says 15-30m
- 13% says 2-5m
- 8% says <2m
4 mins critical path, 1h30m wall clock time (bunch of optional tests)
was around 40 mins on the critical path at the start of the year. gains came mostly from sharding tests, replacing webpack with vite, eslint/prettier with oxlint/oxfmt, caching node modules
My experience has made me more interested in seperating deterministic checks the judegment based reviews.Authentication changes have also taught me that passing inidividual checks doesnt mean an integration works end to end.
Trigger a review agent for any commit which has an open PR and has passed build. Use open weights, or at minimum a different model family from the one which wrote the code. I will be publishing a write up soon(tm), inspired by but different from, cloudflare's post.
I would napkin math the averaged cost per-review to be less than $1, which is less than our most rigorous validation pathway in CI (needs GPU for up to an hour). Time-wise around ~10-15m, with variability - and a bug in Argo with pod scheduling delays. This is fine because developers check in on their own PRs when they do and then make the changes that are needed. A key point here is that we are keeping humans in the loop in the PR.
This is a large monorepo, I would expect much lower numbers for polyrepo setups.
That's impressive! Would certainly read it when it comes out.
teaser question, what parts of your agent PR process can you move to scripts?