CASEL lowers harmful tool-call trajectories in a controlled agent simulation
Original title: CASEL: Consequence-Aware Selective Execution for Reliable Long-Horizon Tool-Using Agents
Harmful trajectories fell from 14.75 percent to 7.02 percent versus an uncertainty-ranked gate, under equal compute.
CASEL is a model-agnostic runtime policy that scores failure probability, consequence, reversibility, ambiguity, and evidence freshness before each tool call. It chooses among direct execution, verification, sandboxing, clarification, approval, or abstention. In a 20-seed simulation with 120,000 test actions per method, mean normalized harm fell from 0.184 to 0.123. The authors say intervention effectiveness was assumed, so the ranking does not prove deployment performance.
Unconfirmed: Results rank allocation rules under simulator assumptions and do not establish deployment effectiveness.
- A probability-times-consequence gate reached 5.82 percent harmful trajectories but used 45 percent more approvals.
- Removing reversibility-aware routing raised the harmful rate to 15.53 percent.
- A trajectory-level risk budget increased harm to 11.82 percent at equal cost.
From the original article
1 min read at arxiv.orgTool-using language agents increasingly modify external systems, yet most runtime safeguards allocate scrutiny by generic uncertainty scores or coarse mutation heuristics.
By Yanyan Zhang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.