Folding an agent's trace into typed run state answers monitoring questions at 0.86 accuracy where raw-trace reading gets 0.48

arXiv 2609.01466·high signalXBlueskyLinkedInCopy link

A live trace model keeps an append-only event ledger, folds it incrementally into typed run state, and compiles per-consumer views. For a human observer's monitoring questions the compiled view used roughly 14-15x fewer input tokens at 5-7x lower cost than a budget-capped single-call read of the raw trace, and scored 0.85-0.87 accuracy versus 0.48. On 120-link sequential-dependency tasks, keeping the running statistic in per-step state succeeded 30/30 where full-context prompting managed 8/30. The authors are explicit that a prompt-level scratchpad matches the fold's accuracy more cheaply, and the fold's remaining edge is deterministic auditability.