Agent Accountability

The property that lets an organization answer, for any action an AI agent took, four questions: *what* did it do, *why*, *on whose authority*, and *who is responsible for the outcome* — backed by evidence strong enough to survive a dispute. Accountability is distinct from its neighbors: transparency is seeing what an agent does, explainability is understanding why, auditability is reconstructing it after the fact — accountability adds a *locus of responsibility*, a named human or organization who owns the consequences. Autonomy erodes it through the classic 'many-hands' problem: an agent's action is the product of the model provider, the framework, the deployer, the prompt author, and the person who granted its scopes, and each can plausibly disclaim their part. Restoring accountability requires four things together — a tamper-evident audit trail, a provable Agent Identity so actions attribute correctly, a named accountable owner, and Non-Repudiation so no one can deny what happened. It is the fourth governance question after 'is it real?' (Agent Washing), 'who is it?' (Agent Identity), and 'did we authorize it?' (Shadow AI).