Agentic Process Governance
An allow is not discharge. The institution, not the agent, pins the goal and the floor. Governance is the relation among the goal, the trace, the proof boundary, and the trust that remains.
This page is the abstract. The PDF is the paper. This is the paper that joins the research to the Compliance Operator. Terms are in the glossary.
Abstract
When an agent may take a different path to the same goal, the word governance is asked to cover too many different things: a log, a person placed on every step, a label of assurance, a check before an action, and a file that has passed a familiar sequence of review steps. Those are real pieces of discrete work. They are not the same object.
This paper states the governed object. Governance is the maintained relationship among the governing goal, the local obligations the goal depends on, the trace of gated actions, the execution boundary, the proof-support set, and the trust that remains. Write P(G_a, x_i | A_i) for the claim that this trace satisfies this goal relative to the execution boundary A_i. The proof inside discharge is the different claim P(G_a, x_i | A_g^i), relative to the proof-support set. The same goal may be met by different traces. Two cases that look like the same journey can still be different claims. Agreement of those obligations on a shared fact is not part of that claim (§7).
The institution pins the goal and the mandatory floor before the run: it fixes and versions them, and the agent cannot edit them. A further requirement, stated here and not derived, is that the party that selects the path does not select the floor. The production of evidence and the gate's decision stay separate. A person who prepares a document has traditionally stood apart from the role that accepts it. An agent can call a tool and receive, in the same run, the evidence the case will rely on. The tool's return does not settle the gate. An *allow*, the gate's decision to permit an action before it executes, is a control on one action. It is not, by itself, the assurance claim. Discharge asks for the proof, coverage of the floor, an acceptable residue, and a support that sits inside the execution boundary. One pinned policy is compiled twice, into the machine that allows or denies a consequential action and into the skill the model reads when it proposes what to do next. That double compilation is a constraint this paper states and does not establish. The mortgage application is the illustration. A fall in income can leave the outcome with a person after the predicates a file can prove have been proved.
The contribution is the separation of an allow from discharge, and the requirement that someone other than the agent pin the floor and the acceptability of residue.
The paper does not supply the rule that chooses between two incomparable sets of assumptions that both cover the floor, and it does not define how two local proofs are shown to use the same shared fact. It does not certify a lending system, and it does not claim that a control has been accepted by an examiner.
Keywords: agentic process; governance; proof boundary; trust residue; goal-relative assurance; delegated authority
Betts, R., Riley, M. and Riley-Betts, J. (2026) ‘Agentic Process Governance’. Riley Betts Ltd. Available at: https://rileybetts.ai/papers/agentic-process-governance