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.
Debugging AI generated code and agent workflows: the failure modes that pass review, pass tests, and break in production.
MicroVMs
The question is not whether an autonomous agent will do something destructive. It is what the destruction can reach when it does.
AI
Green CI, clean review, broken production. The gap is that tests check whether code matches the implementation and production checks whether it matches reality.
AI
After reviewing a lot of agent generated pull requests, the same six failure patterns keep appearing. They share a structure: locally plausible, globally wrong.
AI
Each agent works fine alone. Chain them and quality collapses. The failure is almost always at the boundary between them.
AI
Every mitigation at the model layer can be worked around. The fix is architectural: assume the model will be compromised and limit what it can do.
AI
You cannot assert equality on a model response. You can assert a lot of other things, and most teams assert none of them.
AI
It was sharp for an hour and now it is contradicting itself and reintroducing bugs you already fixed. That is context degradation, and it is manageable.
AI
Your tool is registered, the agent knows it exists, and nothing happens. Here is how to see what is actually crossing the wire.
AI
An honest inventory after a year of daily use: what earns its place in the loop, what I stopped opening, and where these tools are still bad.
AI
Large diffs used to signal that somebody worked hard. Now they signal that a prompt was short. Review has to change shape.
AI
Most agent instruction files are a list of vibes. Here is what belongs in one, based on which rules measurably changed the output.
AI
Models invent plausible package names. Attackers register them. Here is why this attack works and the two checks that stop it.