Original title: Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers
Egor Pakhomov, Erik Nijkamp · September 2026
When artificial intelligence programs run long tasks, their record of events gets too big. This history quickly overwhelms human observers who need to monitor the run. It also overflows the agent's own memory limit, since the agent must read its past steps to decide what to do next. To fix this, researchers built a live traceA complete log of all actions and events during an agent's operation. model. It logs new events into an append-only list, folds them step by step into organized summaries, and compiles special views for each user. For monitoring, an AI reading this compiled viewA pre-processed summary of the trace designed to efficiently answer specific types of questions. answered questions with higher accuracy (0.85-0.87 versus 0.48), while using approximately 14x and 15x fewer input tokens and costing 5-7x less than reading raw traces. For the agent running long tasks, keeping running statistics in step-by-step state succeeded 30/30 times, whereas reading full history worked only 8/30 times. The system also gives a shared, verifiable record for humans.
Imagine you are assembling a complex model kit, like a car or a plane. As you follow the instructions, you add more and more pieces, and the manual gets thicker with steps. If you lose your place, you might have to re-read dozens of pages to find where you left off. In this paper, the agent's trace is like the growing manual of steps taken, and the live trace model is like having a bookmark that keeps track of your progress and lets you quickly check what you've done without flipping through everything.
When AI agents work on long tasks, their action logs can become unwieldy for both human supervisors and the agents themselves. This model makes it easier and cheaper to monitor agents accurately, which helps developers build more reliable AI systems.
The abstract notes that the token and cost reductions are conditional on the schema covering the monitoring questions, so it may not work if questions fall outside the schema. It also identifies a task family where the fold ceases to help, specifically order-sensitive tasks. The evaluation uses an LLM reader as a proxy for human observers, which may not perfectly represent real human performance.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers — Egor Pakhomov, Erik Nijkamp
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.