All posts
Everything published here, newest first.
316 posts · showing 81–90
2026
10 posts-
Your 95% Step Accuracy Is a 60% Agent
Compounding explains part of the gap between a green eval dashboard and a broken agent. The rest is that per-step scoring counts a step as correct when it succeeds on a premise that was already wrong.
-
Four Tool Calls to One Answer: A Research Run, Annotated
A research agent answers one question in four steps. Reading what changed in the transcript between each step explains more about how agents work than any diagram of the loop.
-
Read-Only Was Never the Same as Safe
SELECT-only credentials close exactly one of the four ways an analytics agent can hurt you. The other three don't require the ability to write anything to the database at all.
-
Should Your Agent Remember That, or Just Look It Up Again?
Two questions sort every fact an agent encounters into "write it down" or "fetch it fresh" — and getting the sort wrong is how an assistant ends up confidently quoting a meeting that moved last week.
-
The First Hour of an Agent Incident
Nothing crashed, no alert fired, and a customer just forwarded a policy answer your support agent invented. The hardest question in the first hour isn't what broke — it's when it started.
-
The Supervisor That Should Have Been a For-Loop
Every night the same plan, a model call per dispatch, and about once a week it checked eleven services instead of twelve. A for-loop cannot skip an item; an agent orchestrating one eventually will.
-
Handing an Agent a PDF: Three Paths Into the Context
A PDF never enters a context window — something derived from it does. The counterparty name right and the date wrong usually means the document took the wrong path in.
-
Was It Retrieval or Reasoning? Diagnosing a Wrong Answer
A day spent rewriting the prompt for a wrong policy answer, when the policy document was never retrieved. Three checks in the right order, and one log field that separates "ignored it" from "never saw it."
-
Designing an Agent You Can Redirect Mid-Run
Twelve files into forty, you realize the approach is wrong. Let it finish or kill it — both bad, and the third option had to be designed in before the run started.
-
The Screening Agent That Could Only Be Right
Every available metric said it was working — and every one of them was computed only from the candidates it advanced. The decision destroys the evidence needed to evaluate the decision.