Permission to Retry: authorization and recovery in AI agent workflows

Evidence record

Permission to Retry: authorization and recovery in AI agent workflows

Read the original source ↗

Claim-level citations (2)

• supportsPermission to Retry shows why a saved agent plan cannot authorize another effect: In the 336 controlled recovery histories, the retained-ledger method passed all 42 cases on each native host: 21 confirmed completions, 15 justified denials and six ambiguous outcomes kept reserved rather than retried.

  Abstract, paired host outcomes and retained-ledger policy

• supportsPermission to Retry shows why a saved agent plan cannot authorize another effect: The study executes 336 prospectively frozen synthetic histories across LangGraph and Temporal. Per host, a retained-effect-ledger policy passes 42 of 42, while cached retry, recheck, and stable-key approaches pass 30, 36 and 31 respectively.

  Abstract; frozen 336-history campaign; policy results per host