Autonomous AI agents require task specifications that define not only what to achieve, but what constitutes acceptable achievement. Many agent frameworks treat success as Boolean goal satisfaction or implicit evaluator heuristics, which complicates verification and weakens auditability when multiple execution paths are possible. This paper introduces MANDATE (Multi-Agent Nominal Decomposition for Autonomous Task Execution), a specification framework that adapts tolerance/threshold-based requirements practices from systems engineering to produce variance-tolerant task specifications. MANDATE's central construct is an anchor for a minimum/target/constraints tuple that defines acceptable success and is shared across multiple courses of action (COAs). The framework produces either (1) mandate-as-code, a machine-readable specification containing COAs, risk metadata aligned to the NIST AI Risk Management Framework, and hash-linked decision provenance; or (2) a Gap Analysis Report that identifies precisely which required knowledge or thresholds are missing to complete the specification. MANDATE therefore functions both as a specification methodology and as an automation-readiness diagnostic, while remaining execution-agnostic: downstream systems may execute, govern, and monitor against the produced artifacts.
Elias Calboreanu (2026) studied this question.