The Eval Wasn't About the Model
The useful signal was not which agent looked best on visible tests. It was which one kept Archibus's boring rules intact after the hidden checks showed up.
Agent evals Read issue
Topic file / 2 issues
AI agents
Filed Here
2 public issues.
Agent Swarms Still Obey Amdahl's Law
Parallel agents are useful when the work is actually parallel. They do not remove the serial review, merge, and judgment path.
Agent systems Read issue