02.EAgentic AI Engineering

An agent that can prove why it acted.

The demonstration is the easy half. Whether the agent is still running a year later turns on whether anyone can account for a single action it took.

APRA CPS 230, every action replayable
Sheet 02, the five in full

The problem

The demonstration works and the audit does not

An agent takes an action, the action is correct, and everybody is impressed. Six weeks later it takes one that is wrong, and the only record is a prompt, a model version that has since been updated, and a log line saying the tool was called. Nobody can say what the agent knew, what it was permitted to do, or why it chose that. The response is to turn it off, which is the outcome the work was meant to avoid.

APRA CPS 230 treats a material service failure as an operational risk event, and expects the entity to explain the control that failed. "The model decided" is not a control.

What the work produces

Three things you keep

  1. 01

    A record per action

    What the agent was asked, what it read, which tools it was permitted to call, and what it did. Written as the action happens, not assembled from logs afterward.

  2. 02

    Boundaries the agent cannot argue with

    Permissions enforced outside the model rather than requested inside the prompt. An instruction is a preference. A boundary is a control.

  3. 03

    A replayable trace

    Any past action can be replayed against the inputs it actually had. Without this an investigation is an opinion about what probably happened.

The governed spine

How it answers the spine

An agent acting on a term it has inferred rather than been given is the definition problem with an execution engine attached. The register is what the agent is grounded on, and the decision record is what makes its actions reviewable.

Three artefacts, every engagement

  • The written record of what a contested term means, who owns it, and the date the wording was agreed.
  • Every hop a number takes from the system that first recorded it to the report that shows it.
  • When a definition changed and why, written at the point the call was made rather than reconstructed afterward.
Sheet 03, the method in full

02.ECommission the work

Bring the term two systems report differently.

A 30-minute diagnostic, not a sales call. Bring the disagreement and the deadline that made it matter.

Book a 30-minute diagnosticSheet 04, other ways to reach us