The paper presents a live trace model that incrementally folds an agent's event stream into typed state, then compiles per-consumer views for both the agent and a human observer. For the observer, an LLM reader using the compiled view achieved 0.85-0.87 accuracy using approximately 14-15x fewer tokens and at 5-7x lower cost compared to reading the raw trace (0.48 accuracy). For the agent on 120-link sequential-dependency tasks, mechanisms using the incremental fold succeeded in 30/30 cases versus 8/30 for full-context prompting. The approach enables deterministic auditability and serves both consumers from the same state.
Typed trace folding helps agents and observers manage long runs
An append-only event ledger compiled into consumer-specific views reduced monitoring costs and preserved key state across 120-step agent tasks.
Industry
Egor Pakhomov · Erik Nijkamp
Salesforce AI Research
Research Digest··2 min read
In one line
A single incremental fold of a live agent trace serves both agent and observer with higher accuracy and lower cost than raw trace reading.
Before you read this
Three ideas this paper builds oncontext compactionCompressing or summarizing an LLM's past context into a smaller representation to extend memory and improve reasoning within limited token budgets.71 papers →typed state representationTracking state with explicit data types, separated from21 papers →runtime monitoringAudits an AI agent's live actions, internal states, or effects against constraints or policies to catch errors and attacks.107 papers →
Read it as
Pakhomov and Nijkamp built a live trace system that incrementally converts an agent’s event history into typed state and separate views for agents and human observers.
The authors implemented an append-only event ledger that is incrementally folded into structured run state, then compiled into views tailored to different consumers.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.