In our earlier testing, we found Claude Sonnet 4.5 exhibited context anxiety strongly enough that compaction alone wasn't sufficient to enable strong long task performance, so context resets became essential to the harness design. ...
Context resets were a key unlock: the harness used Sonnet 4.5, which exhibited the "context anxiety" tendency mentioned earlier. Creating a harness that worked well across context resets was key to keeping the model on task. Opus 4.5 largely removed that behavior on its own, so I was able to drop context resets from this harness entirely.
Anthropic Engineering: Harness Design for Long-Running Apps
TIL that compaction behaves differently to a full reset (after which the agent presumably reads in a summary file to get up to speed) with respect to context anxiety... but that explains why compacted sessions always felt like they were performing worse than fresh ones.