Swift

Thread by @jacobtechtavern 3

**Jacob Bartlett** @jacobtechtavern [2026-04-20](https://x.com/jacobtechtavern/status/2046304903362973871)Performance profiling is ridiculously easy these days. Just throw an instruments trace at an agent, and tell it to find your biggest bottlenecks.If agents are good at anything, it’s crunching a ton of data in a systematic way. They will spend 30 mins parsing out your gigantic instruments trace and automagically finding out why your screen is so slow.ImageImage
**Temel Günaydın** @gunaydin\_temel [2026-04-20](https://x.com/gunaydin_temel/status/2046331188101460076)Is it claude?
**Jacob Bartlett** @jacobtechtavern [2026-04-20](https://x.com/jacobtechtavern/status/2046331292682170539)I like to use Claude and Codex.
**Tim Davison ᯅ** @timd\_ca [2026-04-20](https://x.com/timd_ca/status/2046322690307596330)I've been dreaming about doing this with Metal GPU captures for debugging compute kernels. I got pretty far!! but I'm stuck decoding some of the buffers :/ it replays a lot of the capture... very fragile.
**John Marc** @ioloro [2026-04-20](https://x.com/ioloro/status/2046309049218760825)100000% I ended up making a skill for it. Big thing was making JSON output from the trace file to save on context.