Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers

TL;DR - This paper proposes a live trace model that incrementally folds long-horizon agent events into typed state and generates compact views for agents and human observers. It improves monitoring efficiency and accuracy while providing deterministic, auditable state management.

• Observer views used roughly 14–15× fewer input tokens and cost 5–7× less than budget-capped raw-trace reading, while improving accuracy from 0.48 to 0.85–0.87.
• On 120-link sequential-dependency tasks, per-step running-state mechanisms completed 30/30 trials versus 8/30 for full-context prompting.
• A prompt-level scratchpad matched the fold’s task accuracy at lower cost; the fold’s added benefits are deterministic auditability and shared state for observer views.
• The authors identify eleven candidate trace-folding requirements and demonstrate an order-sensitive task family where folding no longer helps.