V excited to join @jonas and the incredible @iterate team to build actually-useful AI agents for startup ops. Time to add my voice to the very quiet, completely underrepresented AI conversation on X.
Turns out we were awaiting RPC calls + DB writes for every response chunk for every LLM call in our internal product. Ended up just about fast enough to not be suspicious of doing something so bone-headed
Put revhub.sh at the beginning of any github pull request URL and get a better way of reviewing:
- wrapper for diffshub by @pierrecomputer so fast virtualised rendering + file tree
- quickly view changes since a commit, or since a review
- group based on filepath
Putting capnweb/RPC + codemode at the heart of everything makes it super easy to test slow, expensive unreliable stuff like LLM requests deterministic, reliable and fast
This is our full system, with an "AI-like" responder defined in and scoped to a single test.
In the next version of middlewright - support page popups, as always, with natural playwright code with zero changes.
Your agent just does what it does normally - page.waitForEvent('popup') - and it just works. Popup video gets overlayed in the rendered output
