Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers
A live trace model is presented as an append-only event ledger folded incrementally into typed run state and compiled into per-consumer views. For the observer side, evaluated with an LLM reader as proxy, the compiled view answers monitoring questions using approximately 14x and 15x fewer input tokens (by reader) and at 5-7x lower cost than a budget-capped single-call reading of the raw trace, with higher accuracy (0.85-0.87 versus 0.48). For the agent, on 120-link sequential-dependency tasks, mechanisms that maintain the task's running statistic in per-step state succeed where full-context prompting fails (30/30 versus 8/30 under a clean protocol, n=30).
A research paper introduces a live trace model for long-horizon agents, using an append-only event ledger folded into typed run state and compiled into per-consumer views. Evaluations show significant token and cost reductions for human observers and improved success rates for agents on sequential-dependency tasks.
The live trace model uses an append-only event ledger with incremental folding into typed run state, enabling per-consumer compiled views. For observers, an LLM reader proxy achieved 14-15x token reduction and 5-7x cost reduction with higher accuracy (0.85-0.87 vs 0.48) compared to raw trace reading. For agents, maintaining running statistics in per-step state succeeded on 30/30 sequential-dependency tasks versus 8/30 for full-context prompting, indicating that structured state management outperforms raw context for long-horizon reasoning.
This approach addresses scalability challenges in monitoring and operating long-horizon autonomous agents, which are increasingly deployed in enterprise and developer workflows. The token and cost reductions suggest potential for more efficient human oversight and lower operational costs, while the agent-side improvements indicate a path toward more reliable long-duration agent execution.
The model can reduce monitoring costs and improve accuracy for human oversight of AI agents, potentially lowering operational expenses and increasing trust in autonomous systems. For agent developers, the stateful approach may enable more complex long-horizon tasks with higher success rates, expanding the scope of deployable agent applications.
Next observable signals include adoption of similar trace models in agent frameworks, further benchmarks comparing stateful vs. context-based agent architectures, and integration of compiled views into monitoring dashboards. The co-design of questions and schema suggests a need for standardized trace schemas to enable broader applicability.