Making AI agents accountable through tracked decision histories

Current agentic AI frameworks lack quantifiable provenance to assign responsibility when complex multi-agent compositions cause harm. This work argues that explicit provenance across the full agentic lifecycle—from task planning through execution—is structurally necessary, not optional. The authors formalize responsibility through causal attribution functions and a responsibility tensor, demonstrate that provenance is computable and interventionable online across four lifecycle layers, and ground the framework in a concrete agentic incident. The contribution is primarily conceptual and architectural rather than a new algorithm; it identifies gaps in how responsibility is tracked and proposes a framework for making it enforceable.