Glossary: Terms for Governing AI Agents
Plain definitions of the terms this field runs on — each with a short explanation, FAQs, and links to the pages that go deeper.
Agentic AI Governance
The discipline of controlling autonomous AI agents so their actions are authorized, attributable, reversible where necessary, and evidenced end to end — the agent-layer complement to model governance.
AI Agent Audit Trail
An AI agent audit trail is the record of every tool invocation, system change, policy decision and human approval an agent made — attributable to one agent identity, reconstructable months later.
AI Agent Cost Observability
Attributing every unit of agent spend to the agent, workflow and owning team, and capping it with limits that hold under load — measured as cost per completed case, not cost per token.
AI Agent Kill Switch
A kill switch stops an agent including in-flight work, works independently of the agent code, preserves a full action record, and blocks restart without a different approver — proven by a five-step drill.
AI Agent Orchestration
Orchestration coordinates multiple agents or multi-step runs: routing, context passing, failure handling. Governance must hold at every hop with per-agent identity and capped delegation.
AI Agent Sandboxing
AI agent sandboxing limits what an agent can reach — systems, data, tools — so ungated actions stay harmless. The permission boundary lives in infrastructure the agent cannot rewrite.
AI Control Plane
The governance layer outside AI agents: registry, per-agent identity, policy gateway, approval gates, evidence trails and stop controls. It decides what agents may do, before they do it.
AI-DLC
AI-DLC is a governed AI development lifecycle that extends software delivery practices to autonomous agents with cost observability, approval gates and evidence tracking.
AI Gateway
The enforcement point between agents and the models/tools they call: verifies identity, checks policy before execution, enforces budgets, writes evidence. The agent credentials only work through it.
AI Model Risk Management
The lifecycle discipline for AI model risk — validation, monitoring, documentation, accountability — anchored in US banking by SR 11-7 and SR 26-2, now stretching to agentic AI.
Agent Identity Propagation
Carrying a distinct verifiable identity for each agent through every call — models, tools, downstream systems — so every action is attributable, with capped delegation across agent chains.
Agent Observability
Traces, metrics and guardrail telemetry making agent behavior inspectable — the visibility half of governance. It records what happened; it does not decide what may happen.
Agent Registry
The authoritative inventory of every AI agent: purpose, owner, access, approval and production version. It answers "what exactly is running?" with deployment truth, not last-quarter documentation.
Approval Gate
An approval gate pauses an AI agent before a consequential action so a named human can approve, reject or return it — with the proposal, evidence and reason recorded.
Failure Drill
A timed 45-minute rehearsal: stop the agent, verify zero post-stop actions, list everything it wrote, confirm rollback paths, prove restart requires approval. Governance claims, tested.
Forward-Deployed Engineering
FDE puts the engineers who build the system inside the customer environment and context. For AI agents, building against the workflow reality beats building against a spec.
Governed Agentic Workflows
Business processes where agents handle repetitive work, every consequential decision passes policy or a named human, and every step leaves evidence. The workflow is the unit of governance.
Governed AI Agents
Governed AI agents carry their own identity, act within scoped permissions, pause for human approval on consequential actions, and leave a record of everything they did.
Human-in-the-Loop AI
Human-in-the-loop AI prepares and proposes; the human decides. A designed interface of approval gates, evidence and review — not a rubber stamp.
MCP Governance
MCP governance decides which agent may connect to which MCP server, which tools it may invoke, what credentials it uses, and what evidence each call leaves — six pillars from identity to inventory.
Working through these terms for a live deployment?
The Learn hub has the longer guides, and the agent readiness checklist turns the vocabulary into 36 checks.