Controlling agent actions: a reference architecture
A vendor-neutral reference architecture for deciding, enforcing and recording what AI agents may do, with a maturity model and a mapping to the EU AI Act, ISO/IEC 42001 and NIST AI RMF.
Key findings
- A single enforcement point in front of every tool call is the control that most other controls depend on; without it, logging and approvals have nothing reliable to attach to.
- Agents should hold their own identities with task-scoped, short-lived credentials. Sharing a human user's session makes actions unattributable and permissions impossible to narrow.
- Tracking whether untrusted content influenced a request, and tightening policy when it did, reduces exposure to prompt injection more than input filtering alone.
- The same decision log serves engineering debugging, security investigation and regulatory record-keeping, provided it is structured and versioned from day one.
Contents
Executive summary
Organisations are moving from chat assistants to agents that act: they file tickets, change records, issue refunds and run code. The governance question changes with them, from what may the model say to what may the system do.
This paper proposes a reference architecture built around three responsibilities: decide (policy), enforce (a single gate in front of every tool) and record (a structured decision log). It then sets out a four-level maturity model and maps the architecture to the main regulatory and standards frameworks.
1. Scope and method
The architecture applies to any system in which a model can trigger side effects through tools, whether built on a vendor platform, an open-source framework or custom code. It was derived from [describe method: interviews, deployments reviewed, documentation analysed] and is deliberately technology-neutral.
2. The architecture
2.1 Decide
Policies are explicit, versioned rules over a structured request: principal (agent and the user it acts for), action (tool), resource (target) and context (amount, data classification, whether untrusted content is present). Policies live in a repository, are reviewed like code and are tested against a table of expected decisions.
2.2 Enforce
A single enforcement point sits on the only path from agent to tool. It obeys four outcomes: allow, allow with limits, escalate to a human, deny. Credentials for tools are held by the enforcement point, not by the agent, so bypassing the gate means having no credentials.
2.3 Record
Every request and decision is written to an append-only log with the policy version, the reason and, for escalations, the reviewer’s identity and response. The log is the source for debugging, incident response and compliance evidence.
3. Maturity model
| Level | Name | Characteristics | Typical gap |
|---|---|---|---|
| 0 | Implicit | Agent holds broad credentials; behaviour constrained only by the prompt | No enforcement, no record |
| 1 | Allowlisted | Fixed list of tools; basic logging | Arguments and context unchecked |
| 2 | Policy-gated | Every call evaluated by versioned policy; escalation for high-risk actions | Policies not tested; taint not tracked |
| 3 | Governed | Tested policies, per-agent identity, taint tracking, circuit breakers, audit-ready logs | Continuous evaluation and red teaming |
4. Mapping to frameworks
| Architecture element | EU AI Act (high-risk systems) | ISO/IEC 42001 | NIST AI RMF |
|---|---|---|---|
| Versioned policies | Risk management system; technical documentation | AI risk treatment; operational planning and control | Govern; Manage |
| Enforcement point | Accuracy, robustness and cybersecurity | Operational controls | Manage |
| Human escalation | Human oversight | Controls for human oversight | Manage |
| Decision log | Record-keeping (automatic logging) | Documented information; monitoring | Measure; Manage |
This mapping is indicative and is not legal advice. Check obligations against the current texts and guidance.
5. Recommendations
- Introduce a single enforcement point before adding more tools.
- Issue each agent its own identity and remove shared human credentials.
- Track taint from untrusted inputs and make policy depend on it.
- Version policies and logs together so any past decision can be explained.
- Review escalation rates monthly: too many and reviewers stop reading; too few and the thresholds are probably wrong.
Appendix: glossary
Agent. A system in which a model selects and triggers actions through tools. Enforcement point. The component that applies policy decisions on the path to a tool. Taint. A flag recording that untrusted content influenced a request.