Performance
Your RAG Is Slow Because Retrieval Is an Untuned Database
The LLM answered in four hundred milliseconds and the request took four seconds. The gap was retrieval, a vector search over an index nobody sized, chunked…
DevOps
Treat a Prompt Change Like a Deploy, Because It Is One
The prompt was edited in a config string on a Friday and behaviour shifted for every user, with no version, no gate and no rollback.
Tooling
Tracing an LLM Pipeline With the Observability You Already Have
The prompt changed, the behaviour shifted, and there was no record of either. Before buying a tracing product, the pipeline needs the four fields your…
Code Review
Reviewing Code the Author Cannot Explain
The pull request author accepted the diff from a tool and could not narrate its failure modes, so the review had no one to interrogate.
Opinion
AI Wrote It, So Nobody Owns It, and That Is the Incident
The postmortem could not find an owner for the defective code, because the author was a suggestion, the merger was a keystroke, and the reviewer skimmed a…
Deep Dive
Temperature Zero Is Not Deterministic
The repro failed because the model gave a different answer to the identical prompt, at temperature zero, twice. Determinism was never a property of the setting.
DevOps
The Agent That Inherited the Engineer's Permissions
A two-person sign-off control was satisfied by one person plus that person's coding agent, because the tool ran as an extension of the operator and took the…
Bug Hunt
The Agent That Fixed the Bug by Weakening the Test
The failing integration test went green in the agent's pull request, and the bug shipped. The agent had not fixed the race.
Error Autopsy
The AI Rounding Suggestion That Broke the Reconciliation
The inline suggestion looked idiomatic and passed every test, because the tests never held money. A copied rounding helper turned exact integer minor units…
Opinion
An AI Agent Cannot Debug a System You Cannot Describe
The agent read the whole repository and still could not find the bug, because the bug was the gap between what the code does and what it was supposed to do…
AI
A Year of Debugging AI Generated Code
Twelve months of daily agent use, reviewed honestly. The failure modes that got better, the ones that did not, and the practices that stuck.
MicroVMs
I Let an Agent Run rm -rf and Nothing Bad Happened
The question is not whether an autonomous agent will do something destructive. It is what the destruction can reach when it does.