A map of AI agent governance techniques
Thirteen families of controls for AI agents, where each one sits in the request path, and which risk it actually reduces.
Contents
An AI agent is a model that can take actions: call an API, run code, send an email, move money. Governance is the set of controls that decide which of those actions happen, under whose authority, and with what record. This article maps the main techniques so that later pieces can go deep on each one.
The short version: most real risk comes from the combination of untrusted input, access to sensitive data and the ability to act. Each technique below breaks one link in that chain, and none of them is enough alone.
Where the controls sit
It helps to picture a single request travelling through an agent. Untrusted content comes in, the model plans, a tool call goes out, and a result comes back. Controls attach at each step, and a few cut across all of them.
Controls on the action path
Runtime policy enforcement
A policy gate sits between the agent and its tools and evaluates every proposed call against explicit rules before it runs. It is deterministic, testable and independent of the model’s reasoning, which is why it is the backbone of most serious deployments. The companion piece, Runtime policy enforcement for AI agents, covers implementations in detail.
Sandboxing
When an agent runs code or touches a file system, sandboxing limits the blast radius: an isolated container or micro-VM, a read-only mount, and network egress restricted to an allowlist.
Human in the loop
Some actions should wait for a person. The design question is not whether to ask but what to show: a reviewer who sees “approve tool call?” will click yes; a reviewer who sees the amount, recipient and reason can actually decide.
Controls on inputs and outputs
Prompt injection defence
Any text the agent reads can contain instructions. Filters and classifiers help at the margin, but the durable defences are architectural: keep untrusted content away from privileged tools, and treat model output that was influenced by untrusted content as untrusted too.
Output guardrails
Validate what the agent produces before it reaches a user or another system: schemas for structured output, allowlists for URLs, checks for personal data.
Memory integrity
Long-term memory turns a one-off injection into a persistent one. Record where each memory came from, let users see and delete it, and do not let content from untrusted sources write to memory that privileged steps later read.
Cross-cutting controls
| Control | What it answers | Typical mechanism | Main risk reduced |
|---|---|---|---|
| Agent identity and access | Who is acting, on whose behalf? | Per-agent service identities, OAuth on-behalf-of tokens, short-lived scoped credentials | Over-privileged agents, unattributable actions |
| Containment and circuit breakers | What happens when it goes wrong at speed? | Spend and call budgets, rate limits, anomaly-triggered shutdown | Runaway loops, cost blow-ups, cascading failures |
| Observability and audit | What did it do and why? | Structured traces of prompts, tool calls, policy decisions and approvals | Undetected misuse, failed audits |
| Multi-agent trust | Can this agent trust that one? | Signed delegation, capability narrowing at each hop | Confused deputy across agents |
| Supply chain | Is this tool what it claims to be? | Pinned versions, provenance checks, review of tool descriptions | Malicious or changed tools and MCP servers |
| Evals and red teaming | Does it behave as intended? | Scenario test suites, adversarial testing, regression on each change | Silent behaviour drift |
How the frameworks map onto this
Regulation and standards rarely name techniques; they ask for outcomes. Roughly:
- EU AI Act. For high-risk systems, the Act requires record-keeping (automatic logging), human oversight and technical robustness. Observability, human-in-the-loop and policy controls are how teams evidence those requirements.
- NIST AI RMF. The Govern, Map, Measure, Manage functions are a lifecycle, not a control list. The techniques above live mostly in Manage, with evals in Measure.
- ISO/IEC 42001. A management-system standard: it asks you to identify AI risks, choose controls and show they operate. This map is a candidate control set for that exercise.
Where to start
If you deploy one agent next quarter, the order that buys the most risk reduction for the least effort is usually:
- Give the agent its own identity with the narrowest credentials that work.
- Put a policy gate in front of every tool call, starting with a default-deny allowlist.
- Log every tool call and policy decision in a structured, queryable form.
- Add budgets and a kill switch.
- Add human approval for the small set of actions that are irreversible or expensive.
Everything else, from injection classifiers to multi-agent delegation, builds on these five.