The problem
Long-running agents accumulate tool results, plans, and stale observations. Passing everything forward increases cost and makes the next step less focused.
The Jev question
Score which items are still relevant to the current task and gate what can be dropped. Ask small Jev questions over the current state instead of asking a chat model for a full summary.
The production loop
Compare token savings with task completion and recovery rates. Keep a reversible trace so an agent can restore context when its next step fails.