Parsing the Stream: A Live Trace Model for Long-Horizon Agents and Their Observers
The paper argues that long agent traces need to be folded into structured state, not reread raw.
Pakhomov and Nijkamp describe a live trace model: an append-only event ledger that is incrementally compiled into typed run state and separate views for humans and agents. In their observer tests, the compiled view used far fewer tokens and lower cost than reading the raw trace, while improving accuracy under the reported setup. For the agent side, maintaining per-step state solved sequential dependency tasks that full-context prompting often missed, though a prompt scratchpad matched accuracy at lower cost. The authors frame the fold’s remaining advantage as deterministic auditability and reuse of the same state for monitoring. ArXiv · AI/CL/LG's note
Pakhomov and Nijkamp describe a live trace model: an append-only event ledger that is incrementally compiled into typed run state and separate views for humans and agents. In their observer tests, the compiled view used far fewer tokens and lower cost than reading the raw trace, while improving accuracy under the reported setup. For the agent side, maintaining per-step state solved sequential dependency tasks that full-context prompting often missed, though a prompt scratchpad matched accuracy at lower cost. The authors frame the fold’s remaining advantage as deterministic auditability and reuse of the same state for monitoring. ArXiv · AI/CL/LG's note
score 4