All posts
Everything published here, newest first.
316 posts · showing 51–60
2026
10 posts-
Append or Supersede? What to Do When New Information Contradicts Agent Memory
Most agent memory systems only know how to add. The moment a fact changes, the store holds both versions and the retriever picks whichever one scores higher — which is why a scheduling agent keeps booking a room that was decommissioned last month.
-
Retry, Repair, or Replan? A Recovery Ladder for When an Agent's Step Fails
Blind retry burns money on a call that will never succeed; premature replanning throws away correct work and produces a worse plan. The deciding signal isn't how bad the error looks — it's what the error proves false.
-
Your First Production Agent Should Be Invoice Reconciliation — Because the Ledger Grades It for You
The best first agent isn't the most impressive one — it's the one whose environment tells you when it's wrong. Here's how to score candidate tasks for a built-in oracle, and the two ways that oracle quietly lies to you.
-
Framework or Raw SDK? The Test Is Whether You Can Print the Exact Prompt
The framework debate is usually argued on taste. There's a faster test: build a two-tool agent and try to dump the exact request it sends on turn four. How hard that is predicts your next six months of debugging.
-
Your Eval Ran Each Case Once. That's Why It Missed the Flake.
A forty-case suite can pass every case and still hide a bug that reaches users one week in five. The number missing from most agent evals isn't coverage — it's how many times you ran each case.
-
Agent or Just a Script? Count the Branches That Depend on Reading Something
There is a ten-minute test that settles whether your automation needs an agent loop or just a script with a couple of model calls in the middle — and the two ways of guessing wrong are not equally recoverable.
-
Your Two Reviewer Agents Disagree. Who Breaks the Tie?
Running the same document past two agents is easy. Deciding what happens when they return opposite verdicts is the part most multi-agent designs quietly leave undefined — and the default resolution is usually the worst one available.
-
Cached or Re-Called? Giving Every Agent Tool a Staleness Budget
An agent that reads a value once and acts on it four minutes later isn't reasoning badly — nobody told it how long that value stays true. Per-tool staleness budgets, and the one rule that matters most before a write.
-
One Worker Pool or Two? Separating Fast Agent Runs From Slow Ones
A handful of long research runs can starve hundreds of eight-second ticket replies without a single error firing. The fix isn't more workers — it's deciding, at admission, which lane a run belongs in.
-
Your Agent Retried the Tool Call. The Customer Got Charged Twice.
A timed-out tool call is not a failed one — it is an unknown one. Here is why "just retry" quietly doubles real-world side effects, and where to put the idempotency check so it cannot.