Writing
Notes from production
What actually broke, what we measured, and what we changed. Written by the people teaching the cohorts, from systems they run.
Build the eval harness before you build the feature
Teams that ship reliable LLM systems almost always wrote the measurement first. Here is what that looks like in practice, and why the ordering matters more than the tooling.
Read the article5 articles
Reading a retry storm: the same incident in three graphs
The staging incident looked like a traffic spike on every dashboard we had. Here is what it actually looked like, and which graph would have told us in thirty seconds instead of twenty minutes.
Timeouts, retries, and the agent that took down staging
An agent with a three-second timeout and three retries turned one slow dependency into a thousand requests a minute. The fix was not a smaller number - it was a budget.
Your agent is only as good as its tools
Most agent failures are misread as reasoning failures. Look closer and they are tool design failures - bad granularity, ambiguous errors, and schemas that invite mistakes.
Inference cost is a design problem, not a billing problem
Teams reach for a cheaper model when the real win is upstream: fewer calls, smaller contexts, and caching the things that never needed a model at all.