⚡ Quick Summary
Published by the Digital Transformation Agency, this addendum to the AI technical standard for Australian Government provides best-practice guidance for agencies implementing agentic AI. It states that all statements, criteria and general guidance in the AI technical standard remain applicable, and that agencies exploring, developing or using agentic AI are expected to apply both documents together.
The addendum maps key drivers to lifecycle-specific controls: absence of governance and safeguards; poorly scoped data retention; unclear decision points; weak orchestration and data flow controls; inappropriate agent, model or technology selection; inadequate evaluation; insufficient tool and protocol controls; and absent continuous monitoring. Each driver is answered by a statement (AGT.1 to AGT.8) with its criteria, plus existing Statement 41 and Criterion 147 on decommissioning.
Supporting sections define key terms such as control tower, data exfiltration, observability and the human-in-the-loop, human-on-the-loop and human-out-of-the-loop oversight modes; describe the perceive, reason and act cycle; and set out agent types, an orchestration layer of framework, workflow engine and event bus, and emerging interaction protocols MCP, A2A, ACP and ANP.
🧩 What’s Covered
The document combines a background section with eight statements organised by AI lifecycle phase.
- Key terms: definitions of control tower, data exfiltration, human-in-the-loop, human-on-the-loop, human-out-of-the-loop, human oversight and observability, with existing terms deferred to the key terms page of the parent technical standard.
- Key drivers and best practice guidance: a table pairing each driver with the guidance it produces and the resulting criteria, covering governance and safeguards (AGT.1), memory management (AGT.2), design (AGT.3), data (AGT.4), train (AGT.5), evaluate (AGT.6), integrate (AGT.7) and monitor (AGT.8).
- Applicability: agencies must follow the Policy for the responsible use of AI in government, the AI technical standard, ASD guidelines on careful adoption of agentic AI services, the PSPF, the ISM, the Australian Privacy Principles and Commonwealth legislation including the Public Service Act 1999, Privacy Act 1988 and Archives Act 1983; systems should run under human oversight such as a human-on-the-loop model with auditable reasoning logs.
- Background: the AI agent and agentic AI concepts, the perceive, reason and act cycle with feedback, models, memory, tools and feedback technologies, agent types (planner, executor, researcher, router, supervisor, assistant, monitoring, coordination, communication), orchestration as framework, workflow engine and event bus, and the three oversight modes.
- Statements AGT.1 and AGT.2: governance and safeguards, assignment of agent responsibilities with human accountability, human oversight, explainability and an orchestration layer; memory management covering state, short-term and long-term memory, retention, leakage, poisoning and record keeping.
- Statements AGT.3 to AGT.5: design of agentic workflows (goals, sequencing, guardrails, identity and access, self-correction and fail-safes, tools, decision points, scalability); data readiness as a prerequisite, agent-to-agent data exchange and dynamic routing; selection of agents and models, trade-offs, and prompt engineering techniques including zero-shot, few-shot, role-based, chain-of-thought, prompt chaining and ReAct.
- Statements AGT.6 to AGT.8 and decommission: evaluation of robustness and of agent-tool, agent-environment and agent-memory interaction using ground truth and robustness testing; tool and protocol selection with emerging protocols MCP, A2A, ACP and ANP; continuous monitoring and a control tower; and secure decommissioning under Statement 41 and Criterion 147.
💡 Why it matters?
Agentic AI introduces failure modes that single-model governance does not cover: cascading errors across chained actions, memory that retains too much, tool misuse and injection attacks, and drift that stays invisible without continuous monitoring. The addendum gives agencies and their assurance teams criterion-level checks tied to the lifecycle phases they already use under the AI technical standard, so agentic controls can be assessed alongside existing obligations rather than separately. It also supplies shared vocabulary — orchestration layer, control tower, HITL, HOTL and HOOTL — that helps governance, security and engineering teams agree on what must be logged, approved, monitored and shut down.
❓ What’s Missing
The addendum is deliberately additive: it assumes familiarity with the AI technical standard and repeatedly defers to it instead of restating requirements, so it is not self-contained. It carries no publication date, version number or review cycle, which makes currency hard to judge for fast-moving items such as the interaction protocols it names. Several criteria are written as “may include” lists rather than testable requirements, and no maturity levels, evidence artefacts, assurance methods or worked examples are given. The decommission section reuses existing Statement 41 and Criterion 147 without adding agentic criteria of its own, and its numbering differs from the AGT series.
👥 Best For
Best suited to Australian Government agency teams: AI governance and assurance leads mapping agentic controls against existing obligations, risk and compliance officers assigning accountability and oversight modes, and architects and engineers designing orchestration layers, memory management, tool access, evaluation and monitoring pipelines. Security and operations teams will find the monitoring and decommission criteria useful for defining control tower coverage and shutdown procedures.
📄 Source Details
Agentic AI addendum to the AI technical standard for Australian Government, published by the Digital Transformation Agency. No publication year, version number or edition is printed; the document runs to 32 pages and carries “Official” markings on every page. No authors are named. The only URL-like reference printed is the partial path “key terms | digital.gov.au”, which points to the parent standard rather than to this document. The full 32-page text extraction was available for review.
