From Connectors to Harnesses: Runtime Action Spaces, Forensic Readiness, and the Auditability of Agentic AI

Agentic AI turns model outputs into external action through configurable runtime infrastructure. Connectors expose tools and external services; skills package reusable procedures, instructions, scripts, and resources; harnesses integrate these components with orchestration, permissions, credentials, memory, execution environments, approval policies, and event handling. The paper treats these as nested analytical layers that move from external affordance, through reusable procedure, to coordinated execution.

This progression creates an audit problem. Runtime configuration determines the operations available to an agent, model-visible context conditions the path selected through those operations, and observability settings determine which parts of the executed path survive as evidence. A contested action therefore requires reconstruction of both historical configuration and execution provenance. This places forensic readiness at the center of agent audit and extends it across algorithmic supply chains in which evidence is divided among model providers, agent platforms, skill sources, connector servers, cloud services, and deployers. When the acting stack also controls the principal records used to assess its conduct, accountability becomes evidentially dependent on that stack. The same architecture shapes meaningful human oversight and the value of transparency artefacts. Because runtime reconfiguration can move faster than evidentiary infrastructure, agentic AI also sharpens the pacing problem of technology governance.

The paper derives a reconstruction-oriented governance agenda built around minimum sufficient evidence, independent corroboration, and cross-party access.