glassbox

Glassbox (https://miniglass.vercel.app/) makes the reasoning of autonomous agents visible. In autoresearch loops, agents edit code, run experiments and keep whatever scores better, and every such loop is effectively an RL environment with a task, a mutable state and a verifier handing out reward. As these loops move toward recursive self-improvement, the hard question becomes not whether the agent made progress, but whether it understood something or just gamed the verifier. Glassbox plugs into the agent’s harness: the agent must state a hypothesis and a prediction before every change, and the harness measures the result. In one example, the agent rebuilds a microservice app (HotROD) from its runtime traces alone and is scored on traces it never saw. Every claim is marked ● observed (measured by the harness), ◇ agent (the agent’s own words) or ◆ inferred (computed by Glassbox, including independent SubQ checks that each change tests what it claimed). The result is a replayable map of what the agent believed, what the evidence confirmed, and what was only a guess.