Beyond Tool Success: Outcome-Finality Contracts for Durable Enterprise Agents

Enterprise agents that call external systems face two separate questions: what happened to an operation after an ambiguous failure (effect finality), and whether the business goal has actually been reached (business-outcome finality). This preprint studies both with executable, deterministic models whose state spaces are enumerated exhaustively. In a 16-case retry matrix (prior application × native idempotency × readback × duplicate dispatch), blind retry produced duplicate effects in 12 cases, idempotency-only retry in 6, and reconcile-first in 0; reconcile-first escalated 4 cases to manual review when neither readback nor idempotency could resolve the ambiguity. In an 8-state completion matrix, effect-confirmed completion misclassified 4 states and correction-only completion 2, while an explicit business postcondition (correction observed AND ERP reconciled) classified all 8 correctly by construction. We argue that effect finality and business-outcome finality should be treated as separate correctness layers in durable agent workflows. Results hold within the models; no production systems, latency, or cost are measured.Code and raw results: https://github.com/yasuss/Beyond-Tool-Success-Outcome-Finality-Contracts-for-Durable-Enterprise-Agents