Delegation Contracts for Agentic AI: A Task-Level Framework for Objectives, Context, Constraints, Acceptance Criteria, Authority, and Escalation

As artificial intelligence systems move from generating outputs to executing multi-step workflows, the control problem shifts from response quality to delegated action. This paper proposes the delegation contract, a six-field task-level framework for agentic AI that integrates semantic task specification with operational authority and escalation. A delegation contract specifies objective, context, constraints, acceptance criteria, authority, and escalation. The first four fields define semantic delegation: what the task means and how acceptable completion will be recognized. The final two define operational delegation: what the agent may do and when decision rights must return to another actor. The framework is positioned relative to principal-agent theory, meaningful human control, authenticated delegation, agent infrastructure, authorization architectures, and contemporary agent governance. Its contribution is integrative rather than a claim that the individual elements are new: it treats purpose, completion evidence, delegated action rights, and return-of-control conditions as one coherent task-level control object. The paper also outlines characteristic failure modes, a lightweight machine-readable implementation pattern, risk-adjusted specification depth, and authority attenuation in multi-agent delegation. The central claim is that reliable agentic autonomy requires more than a clear prompt and more than scoped permissions; it requires an explicit specification of what counts as success, what the agent is authorized to do, and when that authority expires.