The biggest burden for agents performing long coding or debugging tasks is the snowballing volume of tool execution logs. 'Agent-Controlled Forgetting,' proposed in a recently published paper (arXiv:2610.10590), addresses this with a highly practical harness structure.

The principle is clear: instead of keeping entire past tool outputs in the prompt, they are replaced with short summary notes, while original data is stored in a restorable archive. It implements a 'reversible forgetting' loop where the agent can query the archive to retrieve previous execution results whenever they are needed again.

In actual debugging tests, it reduced cumulative prompt tokens by about 50% and significantly lowered API costs per run from $4.38 to around $1.28–$1.44. While there is a roughly 17% latency increase due to archive retrieval requests, it seems like a very reasonable trade-off considering the context window pollution and costs for long-running agents. It's definitely a pattern worth noting given the recent focus on agent harness engineering.