✍️ Blog
Notes on agent harnesses — the scaffolding around models that decides whether an autonomous agent thinks or collapses. Context engineering, tool design, verification loops, and lessons from building self-evolving research agents.
Verification Loops: Why Self-Evolving Agents Need Adversarial Critics
An agent that grades its own work grades it too kindly. The single highest-leverage pattern I've found for autonomous research agents is a separate, adversarial verifier whose j...
Read more →
Tools Are the API Your Agent Actually Sees
You can't prompt your way out of a bad tool interface. The set of tools you hand an agent — their names, their granularity, their error messages — shapes its behavior more than ...
Read more →
Context Is the Real Budget: Notes on Engineering Agent Harnesses
The model is rarely the bottleneck. After a year of building self-evolving research agents, the thing that decides whether an agent succeeds or collapses is how its harness spen...
Read more →