#deep-dive
Every post tagged "deep-dive".
107 posts · showing 11–20
-
Your Retriever Returned 12 Chunks. The Agent Only Really Read 3 of Them.
The correct passage was in the prompt and the agent wrote around it anyway. That's not a retrieval bug — it's a placement bug, and there's a one-minute test that tells the two apart before you touch your embeddings.
-
Your Agents Asked 40 Questions This Week. Which Ones Should Have Reached a Human?
Human attention is the one resource in an agent deployment that never shows up on the bill — so agents spend it freely. Here's a three-way routing rule, a threshold you can actually tune, and the two metrics that tell you when your agent is asking too much or too little.
-
Read, Draft, Send: Where to Draw the Permission Line in a Personal Email Agent
Most email agents get permissions wrong by granting them per tool. The safer line runs through the arguments — who the recipient is, how many there are, and whether the thread was already open.
-
What Should Your Agent Forget? Give Every Memory an Expiry Rule
Agent memory systems have a write verb and a recall verb, but no verb for "stop believing this." Here's how to decide at write time what makes each memory stop being true — and why expiry should downgrade a fact rather than delete it.
-
The Worker Agent Reported Success. Where Should the Check Live?
A worker agent's "done" is a sentence, not evidence — and in effectful work the gap between the two is where silent failures live. Three places to put the verification, and the one most teams pick that buys nothing.
-
Retry, Reroute, or Hand Back: Budgeting an Agent's Second Chances
A tool call fails and the loop tries again — five times, forty seconds, and still nothing usable for the user. Retrying is only one of three legal responses, and picking the wrong one turns a fast failure into an expensive one.
-
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.
-
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.
-
Three Ways to Get Valid JSON, and What Each One Costs
The retry loop was working and quietly biased: the tickets needing three attempts weren't a random sample, and their eventual valid output was thinner than the truth.
-
Your Trace Store Will Outgrow Your Model Bill
Storage passed inference cost in three weeks, and the proposed fix — sample at 5% — would have discarded exactly the runs worth keeping. A trace isn't one artifact with one lifetime.