What's this about?
Two papers on agents managing themselves over long runs. ContextPilot teaches proactive context management with fine-grained reinforcement learning — deciding what to keep, drop or summarise *before* the window fills. PILOT in the Loop does live self-improvement for long-horizon agents, adapting during the run rather than between runs (30 upvotes).
Why is proactive context management different from compaction?
Compaction is reactive: the window fills, something must go, and the harness summarises whatever is oldest. Proactive management decides earlier and by relevance rather than age — dropping the failed approach from twenty steps ago while keeping the constraint stated at the start. The Compaction Cliff paper from the 2026-08-28 digest showed exactly why age-based eviction is wrong: the safety rule and the log get summarised at the same rate, and only one of them survives it.
Why does this matter to you?
- Codex's compaction bug this week is the practical version of the same problem — context management is now a billing line, not just a quality one.
- Learning *when* to compact, rather than hard-coding a threshold, is the natural next step and this is the first serious attempt.
- Live self-improvement during a run is a much stronger claim than between-run learning; worth watching whether it replicates.