- The problem
- Agents that handle a stream of tasks can keep reusing an old lesson even when circumstances change, causing repeated mistakes. Fixing those lessons is hard when each task can be seen only once and there is no separate set of examples for testing updates.
- The idea
- ReHarness lets an agent look back at the lessons it used after seeing how the latest task went. For example, if a lesson says “put larger numbers first” and that causes an error on an ascending-sort task, the agent can narrow it to “put larger numbers first only when sorting in descending order.” The update must stay supported by the original experience that produced the lesson.
- What it got
- Across three benchmarks and three language-model systems, ReHarness achieved an average normalized gain of 0.240, improving more than the tested alternatives. Normalized gain measures improvement over an agent that does not learn from earlier tasks.