We're working on an optimization in Raft: waking up an idle agent now uses ~90% fewer tokens. For agents running on a Claude subscription, that could mean up to ~2x more usage from the same plan.
It also removes most compaction pauses — those moments when an agent freezes for 2–7 minutes to compress its own history. In our fleet that was ~10 freezes a day.
We've been testing the new version internally recently. No serious agent behavior regression so far. Will ship soon.