#advanced
Every post tagged "advanced".
48 posts · showing 41–48
-
Migrating Model Versions Without Breaking Your Agent
The new model is better at nearly everything and different at everything, and tool selection is what shifts first. Replay a frozen eval set against both — and put the cost-per-task row in the comparison.
-
Indirect Prompt Injection: When the Tool Result Is the Attack
The attack doesn't come from your user — it comes from the page, ticket, or document your agent fetched. Three conditions have to hold for it to work, and breaking any one is an architectural choice, not a prompt.
-
Designing for Resumability: Agents That Survive a Restart
Serializing the message array is the obvious approach and it produces subtly wrong resumptions. What has to survive a crash is a work ledger, idempotent side effects, and enough state to rebuild context from scratch.
-
Debate or Vote? Getting Real Value From Multiple Model Opinions
Voting harvests independent error; debate harvests challenge. Pick wrong and you multiply your token bill for agreement you'd have had from a single call — and the disagreement, the part worth paying for, gets logged away.
-
Rate Limits Are an Agent Design Constraint, Not an Ops Problem
Backoff assumes a predictable call count. Agents burst, retry at three layers, and generate more traffic the more they're throttled — which makes rate limits an architecture question, not an ops ticket.
-
The LLM Judge That Agreed With Everything
A suite scored 94% and 93% on two versions — but the judge would have passed almost anything. Four biases quietly manufacture those numbers, and one prompt change fixes more than the rest combined.
-
Chain, Blackboard, or Broadcast? Three Ways Agents Share State
Chains lose the evidence, blackboards fight over it, broadcasts drown in it. Pick by which failure your task can absorb — and log provenance from day one or debugging becomes guesswork.
-
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.