Having agents reset context around 225K tokens dropped a g coat by 50-75% for me
retained text
I've been running agents 24/7 and just letting the context fill up. Claude has a setting to compact at a specific time, but I just built it into my harness so that it can happen at the end of a tasks completion instead of randomly. Costs dropped significantly.
document fetched 2026-11-14 00:45 · canonical https://reddit.com/r/ClaudeCode/comments/1xuwp92/having_agents_reset_context_around_225k_tokens