Long-running agents often store intentions that depend on external conditions, like checking a portal for status changes. Each check consumes budget for web access or API calls, and existing systems do not budget checks across competing intentions. The paper introduces BudgetPM, which decides when to check external state under a shared episode budget. BudgetPM-Static uses a lightweight logistic scorer to determine if a check improves the current decision, while BudgetPM-Sequential learns from full-episode hindsight to reserve capacity for future checks. On benchmarks, BudgetPM-Static retains 99.9–100% of unconstrained quality with 42–54% fewer observations, and BudgetPM-Sequential exceeds natural monitoring schedules by 1.92–2.58 Set F1 points under severe scarcity.
Agents can budget external checks without sacrificing task quality
BudgetPM learns when to query external state now and when to reserve limited observation capacity for later intentions.
Academic
Zhengkun Di · Bin Shi · Kai Sun · Yiming Xu · Bo Dong
Xi’an Jiaotong University
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
BudgetPM allocates limited external observations across stored agent intentions to maximize task quality under a hard budget.
Before you read this
Two ideas this paper builds onRead it as
Di and colleagues formulate external-state checking for stored intentions as a resource-allocation problem under a hard episode budget.
The authors study prospective memory, in which an agent stores an intention that should be executed when an external condition becomes true.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.