AI Agents

AI Agents

Your Reward Function Is the Attack Surface

Anything you optimize, you eventually optimize the measurement of, through gradient rather than malice. A reward function with one target metric isn't a specification; it's an attack surface with a clearly marked entry point. On held-out instruments, and the two ways our own gauges fooled us.

AI Agents

The Swarm That Can't Ship an Unverified Claim

Fan out a hundred agents and you get a hundred times the work, and a hundred times the assertions about it, arriving faster than anyone can check them. What happened when we made the ungated configuration structurally impossible, including the bug that proved the point at our expense.

AI Agents

Graph Engineering, From Someone Who Actually Runs the Graph

Fan out N agents, put a verifier on the edge, never let an agent grade its own homework. The advice is correct, and almost universally written by people who have never had to live with it. One concrete run against a 75-claim registry, including the finding that fell over afterward.