Agent Governance · Patterns
Each pattern below is a recurring class of divergence between declared authority and reachable effect — observed across architectures in general, not a claim about any named company or incident.
When approval survives longer than the state it approved. A person approves an action at T1; the state has changed by the time it executes at T2.
When revocation stops the principal but not every derived path. The agent is formally disconnected, but tokens, sessions or delegated chains derived from it can keep running.
When authorized retrieval becomes decision authority. Every document may be permitted, yet the selection and ranking of what is retrieved defines the evidentiary environment the decision is made in.
When an accountable owner cannot necessarily reverse the outcome. An owner is named, but may have no real mechanism to undo or correct the downstream consequence.
When the agent stays the same but its causal reach changes. Identity looks stable while tools, functions, prompts, models or dependencies have changed underneath it.
When multiple agents do not mean independent evidence. Several review agents can appear independent while inheriting the same underlying source of error.
When feedback changes not only behavior, but autonomy. A reward signal can shift not just which action is chosen, but whether human confirmation is still required at all.
When no single agent owns the whole movement. A physical action emerges from the composition of a world model, a planner, a trajectory generator, a force controller and a safety agent.
When authorization survives the handoff but intent does not. Formal permissions pass cleanly between agents, while the original human intent behind them can drift.
A pattern is a hypothesis about where to look. Whether it holds in your workflow is an evidence question.