Keiji Kanazawa describes a personal agent project that grew from summarizing podcasts into scheduled research workflows. After three weeks of agent-directed refactoring, he found misleading abstractions, empty outputs and tests connected to production inferenceAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work..
Keiji Kanazawa restarted with new specifications and compared agent frameworks. He describes using an evaluation rubric to improve video summariesEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria., discovering that an agent had overfit the rubric to one example, and iterating across more videos.
He says he still delegates most code writing and review while setting goals and using adversarial reviews between agentsHuman judgment in AI is the accountable interpretation and decision-making people contribute when setting goals, evaluating evidence, and managing consequences.. His account is a personal experience, not a controlled comparison of the named tools.
Watch on YouTube
