· 6 min read
Agent Evaluation
Docker's Harbor Sandboxes Fix Agent Eval Isolation. They Don't Fix Reward Hacking.
Docker's new Sandboxes integration with the Harbor eval framework runs agent evals in isolated microVMs — a real fix for sandbox escape. But a 16.4% reward-hacking rate from an independent 5-model, 8,989-task study shows isolation and gaming the scorer are two different problems.
Read more