AI Agents

A Production-Readiness Checklist for AI Behavior

Most readiness reviews check whether the service will stay up. An AI feature needs a second pass for what it decides, and for the changes that reach those decisions without a deploy.

Diagnosing an Agent Decision Without an ML Team

Most of what a diagnosis needs isn't model expertise. It's knowing which parts of the explanation were seen, which were said, and which were guessed.

From Trace to Behavior: Instrumenting the Signals That Matter

The Decision-Context Pack: What to Capture Before Something Breaks

An investigation can only use what existed before it started. The way to know what to keep is to run the investigation backwards.

"Why Did It Do That?": A 15-Minute Decision Investigation

One agent decision, a stopwatch, and only the artifacts you already keep.

The Senior-Engineer Tax: What an AI Incident Actually Costs to Diagnose

Four numbers from your last AI incident. The first is the one everyone reaches for, and the least useful of the four.

Triage by Stability: Deciding Which AI Decisions Deserve a Second Look

You can't review every decision an agent makes. Here's how to find the ones actually worth your time.

The Solo-Team AI Reliability Stack: What to Instrument First post

Bad AI decisions don't throw errors. Five layers of instrumentation, in the order they earn their keep, for teams running production AI with no ML engineer.

Evaluating AI Behavior Tools: A Buyer's Checklist

Five questions to put to any tool in a category that hasn't standardized yet, and the one artifact that makes a demo tell you something: an incident you already solved.

Putting a Number on a Bad AI Decision: A Cost-of-Error Worksheet

Six places a bad decision spends money, a worksheet to find as many of them as you actually can, and an honest field for the one you can't.

A Field Guide to Diagnosing Silent Failures in Agents

Five stages, five diagnostic questions. A practical way to figure out where a silent failure actually started, using an incident you already have.

What to Log When Logs Aren't Enough: A Decision-Context Checklist

Every AI decision loses information the moment it's made. Here's a checklist for deciding what's worth preserving before it's gone.

Behavior Tests That Catch What Evals Miss

Your eval suite grades the output. Here's how to build tests that grade the decision instead, using conditions you can construct this week, no new platform required.

Instrumenting the "Why": Capturing Decision Context in Production

A practical, two-tier way to capture why your AI chose what it did — without rebuilding your stack.