CASEL: Consequence-Aware Selective Execution for Reliable Long-Horizon Tool-Using Agents

CASEL: Consequence-Aware Selective Execution for Reliable Long-Horizon Tool-Using Agents

Yanyan Zhang, Fei Shen, Ting Yan and Hao Zhang introduce CASEL, a runtime allocation policy for deciding how to handle each proposed tool call. Instead of relying only on generic uncertainty scores or coarse mutation heuristics, CASEL weighs failure probability, consequence, reversibility, ambiguity and evidence freshness. It chooses among direct execution, verification, sandboxing, clarification, approval and abstention by minimizing a per-mode cost. In a controlled simulation over 20 seeds with 120,000 test actions each, all methods had the same intervention menu and compute cost. CASEL reduced harmful-side-effect trajectories compared with an uncertainty-ranked gate. A gate ranked by the product of failure probability and consequence achieved similar harm, but sent 45 percent more actions to human approval; removing reversibility-aware routing raised the harmful-side-effect rate. The paper also reports that a trajectory-level risk budget increased harm at equal cost. The authors’ argument is that safeguards should allocate scrutiny according to an action’s potential consequences, not uncertainty alone. That matters to their proposal because tool-using agents can modify external systems. The evidence, however, comes from simulation: intervention effectiveness is assumed, not measured, and the results do not establish deployment effectiveness.

Publication: f6547f33-fed4-453e-88c8-bd5335b723fc

attribution — pass: The reading correctly credits the four listed authors with introducing CASEL. Results are framed as findings reported by the paper, not as established facts or Escape Meridian's view.

Authors:Yanyan Zhang, Fei Shen, Ting Yan, Hao Zhang

summary — uncertain: The reading is faithful, but its opening never says CASEL is for AI language-model agents that call software tools. A newcomer meets 'proposed tool call' and 'runtime allocation policy' with no plain context. The AI-agent framing appears only later and indirectly. Smallest repair: in the first sentence, identify CASEL as a safeguard layer for AI language agents whose tool calls can change external systems. Lesser points: 'calibrated' and 'state-dependent' are dropped from the factor list, and the summary omits the headline magnitudes (14.75 to 7.02 percent), though the claims keep them.

Tool-using language agents increasingly modify external systems, yet most runtime safeguards allocate scrutiny by generic uncertainty scores or coarse mutation heuristics.

date — pass: originalPublished is 2026-10-27, which matches the v1 submission date on arXiv. It is not the later feed date of 2026-10-29.

[Submitted on 27 Oct 2026]

support — pass: Each claim matches the abstract: the comparison with the uncertainty-ranked gate, the probability-times-consequence gate sending 45 percent more actions to human approval, the reversibility ablation, and the worse trajectory-level risk budget. The simulation-only caveat is kept prominently. The question follows directly from the stated limitation and does not imply unsupported benefits.

Intervention effectiveness is assumed, not measured: the results rank allocation rules under the simulator's assumptions and do not establish deployment effectiveness.

media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.