Skip to content
Reference architectures
Reference architectureDesign intent

Approval gates for AI agents

A small state machine that turns agent actions into reviewable decisions.

OceanBlu · 4 October 2026 · 4 min read

Problem

Agents become useful when they can act — and risky for the same reason. The gate design below makes every consequential action reviewable without slowing routine work.

States

StateMeaning
ProposedAgent has prepared an action and a decision card
Pending approvalWaiting for the owner or a delegated approver
Approved / EditedA person approved, possibly with changes
RejectedA person declined; the reason is captured for learning
ExecutedAction performed; result attached
Rolled backReversible actions undone on request

The decision card

Every card shows what the agent proposes, the evidence it used, the expected effect, whether the action can be reversed, and what happens if it is declined.

Rules

  • Read and analyse: no gate.
  • Internal write: no gate, always logged.
  • Outward-facing, financial, production-changing: gate every time.
  • Legal: never delegated to an agent.

Why a state machine

It makes behaviour predictable, testable and auditable — the three properties organizations need before they let software act on their behalf.

Start a conversation

Let's explore what comes next.

Tell us about the organization, system or market you are thinking about. We start with questions, not a price list.

Start a conversation