Engineering
How we cut p50 routing latency to 38ms
Connection pooling, speculative fallbacks and a cache that knows when a prompt is deterministic.
Sep 9, 2026 · 7 min read
Postmortem
The 14 minutes evals were not gating deploys
A config flag, a race, and what we changed so it cannot happen the same way twice.
Aug 30, 2026 · 5 min read
Opinion
Your agent does not need more prompts. It needs a trace.
Most "prompt engineering" is debugging without a debugger. Here is what changes when you can see the run.
Aug 14, 2026 · 4 min read