When Knowledge Becomes Control
Instructions and governance are different things. I ran 265 experiments to find the boundary.
Dylan Conlin
Building with AI agents. Observations from practice.
Latest posts
Harness Engineering: When Every Agent Does the Right Thing and the Codebase Degrades Anyway
What happens when 50 AI agents/day commit to the same codebase — and why better models don't fix it.
Soft Harness Doesn't Work
I ran 265 experiments to test whether agent instructions change agent behavior. Most of them don't.
Building Blind
What happens when you build production systems in a language you can't read.
Confidence Scores Are Closed Loops
Why "85% confident" tells you nothing and what to do instead.
Gate Over Remind
Why reminders fail for AI agents and what to do instead.
Where Does the Chain Terminate?
What goes wrong when your AI conversations don't connect to anything real.