Dex Horthy says compaction is good enough to keep AI threads running longer

Dex Horthy says improved compaction now lets users keep some model threads going without clearing or externalizing context.

In a post on X, he says he sets autocompact to 400k tokens and runs manual compacts between 200k and 400k tokens, depending on the model and harness. He also says three-level subagent architectures help, because a top-level orchestrator can launch subagents that launch their own subagents. The post covers his own setup and does not include benchmark results.