Permission to Retry: A Study of Authorization and Recovery in AI Agents (arXiv 2610.31916)

Forum topic · 小凯 · 2026-10-26

Summary

This paper studies execution-time reliability of tool-using AI agents, focusing on how durable agents can resume consequential actions while their authority changes and external outcomes remain uncertain. The authors isolate recovery decisions from language-model behavior, treating checkpoint state, permission to dispatch, evidence of an external effect, and permission to reconcile as distinct obligations. They present a finite conformance study scored against a provider effect ledger hidden from the recovering application. A prospectively frozen campaign executed 336 histories on native LangGraph and Temporal hosts, crossing 14 schedules, three provider contracts, and four matched application policies. All 168 paired host cells produced identical checked outcomes; per host, ledger-based reconciliation passed all 42 histories (21 checked completions, 15 denials, six retained bookings with unknown results), while cached retries, revalidation before retry, and stable-key retries passed 30, 36, and 31 histories respectively. A separate 16-case component study on InfinityOS, an enterprise AI platform, exercises exported permissions, workers, recovery, and settlement code against a repository forge. Findings isolate the roles of current authority, provider contracts, and retained observations in recovery, reporting safety, completion, and unresolved obligations separately per declared application profile.

Paper Overview

Research area: Machine Learning Authors: Deekshith Marla, Kushagra Bhatnagar Published: 2026-10-27 arXiv: 2610.31916

Abstract

A durable agent can resume a consequential action while its authority changes and the action's external outcome remains uncertain. We study execution-time reliability of tool-using AI agents, isolating recovery decisions from language-model behavior. Checkpoint state, permission to dispatch, evidence of an external effect, and permission to reconcile are distinct obligations. We present a finite conformance study that checks these obligations through histories scored against a provider effect ledger hidden from the recovering application.

A prospectively frozen campaign executes 336 histories on native LangGraph and Temporal hosts, crossing 14 schedules, three provider contracts, and four matched application policies. All 168 paired host cells have identical checked outcomes.

Per host, retained-ledger reconciliation passed all 42 histories: 21 checked completions, 15 denials, and six retained bookings with unknown results. Cached retries, revalidation before retry, and stable-key retries passed 30, 36, and 31 histories respectively.

A separate 16-case component study on InfinityOS, an enterprise AI platform, runs exported permissions, workers, recovery, and settlement code against a repository forge. The results isolate the role of current authority, provider contracts, and retained observations in recovery, and report safety, completion, and unresolved obligations separately for each declared application profile.

Key Results

Ledger reconciliation: 42/42 histories passed per host
Revalidation before retry: 36/42
Stable-key retries: 31/42
Cached retries: 30/42

---

#Paper #arXiv #ML