Blog
Why agents drift, and what stops it
Coding agents don't fail randomly. Skipped steps, invented results and instructions that don't survive a long session all come from the same place: nothing outside the model checks whether its own process was followed. These posts explain the mechanism, not just the symptom.
Agent reliability
AI Agents Invent Test Results — Here's Why, and What Actually Catches It
An agent reporting "tests passed" didn't necessarily run them. The mechanism behind fabricated results, and the one property that actually catches it.
Agent reliability
How to Actually Verify What Your AI Coding Agent Did
Five ways to check a coding agent's work, from reading the diff to a hashed-output contract, ranked by what each one verifies and who has to maintain it.
Agent reliability
Why AI Coding Agents Get Worse the Longer the Task Runs
A coding agent that nails a ten-minute fix can drift badly three hours into a migration. The structural reasons long tasks degrade, and what actually bounds it.
Agent reliability
Why Coding Agents Skip Steps — And How to Stop It
Coding agents skip steps because nothing outside the prompt checks whether a step happened. The mechanism behind it, and what actually closes the gap.