All posts
Everything published here, newest first.
316 posts · showing 301–310
2026
10 posts-
Agent Adoption Stalls at Trust, Not Capability
An agent whose errors cluster somewhere describable beats a more accurate one whose errors are scattered. Ask your users when they double-check its work — the answer is your real roadmap.
-
Framework or Plain Loop? Four Questions That Actually Decide It
The core agent loop is under a hundred lines — that's never what a framework saves you. Four situational questions decide it, and the one about 3am debugging should carry the most weight.
-
The Read-Only Agent That Wasn't: Auditing What Your Tools Can Actually Reach
A tool named `check_service_health` woke someone at 3am. Read-only is a claim about effects at the far end of every call, and there's a short audit that either proves it or shows you what you actually deployed.
-
The Triage Agent That Should Have Been a Filter
A routing agent replaced a set of rules and made the common case slower, pricier, and less predictable. The lesson isn't "don't use agents" — it's that judgment is worth paying for only where judgment is required.
-
Assert on the Trajectory, Not the Answer
A final-answer check will happily pass a run that took nineteen turns and touched a tool it should never have called. Five lines of trajectory assertions catch what output tests structurally cannot.
-
Trim, Summarize, or Retrieve? Three Ways to Survive a Long Agent Run
Trimming loses the oldest information, summarizing loses specificity, retrieval loses whatever the query missed. Which loss your run can survive is the whole decision — plus one rule that prevents most long-run failures.
-
Reflection Loops Are Not Free: When Self-Critique Makes the Output Worse
A critique step can only catch what was already derivable from the context — and asked to find problems, a model will find them in passages that were fine. Here's when the pattern earns its doubled token bill.
-
Why Your Agent Stops Too Early: The Missing Termination Signal
"Done" in most agent loops means the model stopped asking for tools — indistinguishable from stuck, confused, or bored. Making completion an explicit, validated claim takes about twenty lines.
-
Your Tool Description Is a Prompt: Rewriting Schemas the Model Can Actually Follow
A three-word tool description sent an inbox agent after the wrong thread. Tool specs live inside the prompt, and five specific rewrites to yours will fix more wrong calls than any system-prompt rule.
-
The Retry Storm Your Dashboard Can't See: When Agent Failures Multiply Instead of Add
Three retry layers, each sensible on its own, compose by multiplication — and a failure at step fourteen costs the whole context, not the failed call. Your per-call dashboard will stay green the entire time.