Planning next moves: improving performance when half your stack is someone else's problem
summary
Defines Time to First Token as the user‑felt latency and a degraded session as any 30‑minute window with a >40s TTFT or cursor overhead >10s; P99 TTFT on bad cohorts hits 57s while HTTP2 is ~9s and cached blobs ~1s. Instrumentation gaps force server/client tracing, so fixes include increasing Redis TTL from 30 minutes to 2 hours, prefetching blobs, and removing duplicate client work (saving ~10s). Goal is to cut degraded sessions toward 1% by June with measurable backend and client changes.
AI-written summary · about
chapters (16)
- 0:00Planning next moves
- 1:11Time to first token
- 3:37Reliability at Cursor
- 5:24Degraded sessions metric
- 6:53User patience threshold
- 8:25Cursor overhead breakdown
- 10:11Agent request flow
- 14:47Missing telemetry impact
- 16:52Agent performance investigation
- 18:16Blob caching HTTP1 issues
- 20:44Optimizing blob caching
- 22:49Eliminating duplicate work
- 23:10Results and challenges
- 24:53Composite metrics limitations
- 26:17Goodhart's law in metrics
- 27:21UI affordances in AI tools
