The FirstHelm glossary defines the terms behind AI agent governance — from autonomous agents and guardrails to the EU AI Act articles that apply to them. Each definition links to a deeper guide where one exists.
AI agent
An autonomous software system that pursues goals by choosing actions with limited human supervision, often using tools, APIs or other agents to complete tasks.
Agentic AI
AI systems that plan and act on their own initiative rather than only responding to prompts. The shift from chatbots to agentic AI is why oversight tooling exists.
AI agent control plane
The governance layer that sits between humans and autonomous agents: it registers agents, logs their actions, enforces constraints and routes risky decisions for approval. Frameworks build agents; a control plane governs them.
Autonomy level
How much freedom an agent has to act without human review, from fully supervised to fully autonomous. FirstHelm treats autonomy as an adjustable dial per agent, not a fixed setting.
Adaptive autonomy
Automatically adjusting an agent's autonomy based on its track record and the risk of the action — earned freedom, withdrawn after failures.
Guardrails
Enforced boundaries on agent behaviour: budget ceilings, forbidden actions, rate limits, time windows and data-access rules evaluated before an action runs.
Constraint
A specific, machine-enforced rule evaluated before each action, returning one of three verdicts: allowed, blocked, or needs approval.
Human-in-the-loop (HITL)
A design pattern where a human decision is required at defined points in an AI workflow — most importantly before consequential actions execute.
Approval workflow
The pause-decide-resume cycle: a risky action halts, a human approves or rejects it, and the agent continues or stops accordingly. Silence is never treated as consent.
Intervention
Taking manual control of a running agent mid-task: pausing it, inspecting its state, overriding an action or stopping it outright.
Agent registry
A single inventory of every AI agent connected to an organisation, its framework, owner, autonomy level and status. You cannot govern what you cannot see.
Activity logging
Recording every action an agent takes, with metadata like action type, description, cost and token usage, in an append-only trail.
Audit trail
The tamper-evident record of agent actions, constraint verdicts, approvals and interventions, exportable as evidence for regulators or auditors.
AI observability
Visibility into what agents are doing in real time — dashboards, alerts, anomaly detection — as distinct from governance controls that change what agents may do.
Multi-agent system
Multiple AI agents working together, often specialised by role. Multi-agent setups multiply the oversight problem: more actors, more interactions, more failure modes.
EU AI Act
The EU regulation on artificial intelligence, phased in from 2025. Its Article 12 requires record-keeping and Article 14 requires human oversight of high-risk AI systems — both are operational duties a control plane helps discharge.
ISO/IEC 42001
The international standard for AI management systems (AIMS). FirstHelm operates as the control plane executing an AIMS's policies at runtime.
NIST AI RMF
The US National Institute of Standards and Technology's AI Risk Management Framework: govern, map, measure and manage AI risk.
AI governance
The policies, controls and accountability structures that keep AI systems safe, legal and aligned with organisational intent.
Runaway agent
An agent whose actions exceed intended limits — overspending, deleting data, sending unauthorised communications — because no pre-execution constraint stopped it. The canonical argument for control planes.