Paper introduces "harness-level forgetting": agent prompt/memory updates can break reliable behavior

omarsar0 · x · 2026-08-21

A paper on continual learning for agent harnesses. Modern agents accumulate experience not in weights but in the harness — prompts, memories, tools, skills, and routing rules. Updating any harness component can break previously reliable behavior with the model untouched, which the paper names harness-level forgetting and provides a way to measure. Its Guarded Harness Evolution separates proposing an update from committing it: a Continual Optimizer drafts candidate harness changes from post-execution feedback, while a Continual Evaluator commits only after checking current improvement, historical retention, and validity. Relative gains exceed 10% (post truncated).

Original post →

More from coding & agent

coding & agent channel →