Terms
Every term in this hub.
This hub collects the runtime control terms in the Prefactor glossary. The policies group is the declarative layer: behaviour policies, agent scope, and the enforcement points where a rule stops being a document and starts being a gate. The guardrails group covers the interception layer that checks inputs and outputs against those rules in real time.
The human control group covers the paths that route decisions to people: human-in-the-loop approval, escalation, and intervention. The honest trade-off runs through every entry: every human checkpoint buys certainty and costs latency, and the right balance depends on what the agent can actually do, not on a universal rule.
The isolation group covers limiting the blast radius when prevention fails: sandboxing, containment strategies, context isolation, and trust boundaries. And the kill switch group is exactly what it sounds like: the mechanisms that stop an agent immediately, without a deploy, when something has gone wrong enough that nuance can wait.
Terms are grouped below from declaration to last resort: policies, then guardrails, then human control, then isolation and containment, then kill switches. Each entry links to its full definition, and terms that pair with an applied guide link onward to it.
Policies & enforcement
Agent Behavior Policy An agent behavior policy is a set of rules that define how an agent should act in specific situations: what it can say, what tools it can use, what data it can access, and when it should escalate. Agent Scope Agent scope defines the boundaries of what an agent is authorised to do, which tools it can call, what data it can access, which systems it can interact with, and under what conditions. AI Policy Library An AI policy library is a centralised repository of governance rules, guidelines, and constraints that can be applied to AI agents. Policy Decision Point A policy decision point evaluates a governance rule against the current context and returns an allow, deny, or conditional decision. Policy Enforcement Point A policy enforcement point is the component in an agent architecture where governance rules are actively applied, blocking, modifying, or flagging agent actions that violate policy. Policy Engine A policy engine is the service that evaluates rules in real time and decides whether an agent action should be allowed, denied, escalated, or logged. Policy-as-Code Policy-as-code is the practice of defining governance rules in machine-readable formats that can be automatically enforced at runtime. Runtime Enforcement Runtime enforcement is the application of security and governance policies while an agent is actively running. Runtime Policy A runtime policy is a governance rule that is evaluated and enforced while an AI agent is actively executing, rather than at deployment time.
Guardrails
AI Guardrail Layer An AI guardrail layer is a dedicated component in an agent architecture that intercepts inputs and outputs to apply safety checks, content filtering, PII detection, and policy enforcement before actions reach their target. Guardrail Bypass A guardrail bypass is any technique that causes an AI agent's safety or governance controls to fail to trigger when they should, through prompt crafting, encoding tricks, indirect instruction, or exploitation of edge cases in the guardrail logic. Guardrails Guardrails are runtime constraints that limit what an AI agent can do, what data it can access, and how it can respond. Output Validation Output validation is the process of checking an agent's response against defined rules before it is returned to the user or passed to a downstream system.
Human control
Agent Approval Workflow An agent approval workflow is a defined process that requires designated reviewers to authorize an agent before it is deployed to production. Agent Escalation Agent escalation is the process of an agent transferring control to a human operator or higher-authority system when it encounters a situation outside its competence, confidence threshold, or policy boundaries. AI Oversight AI oversight is the ongoing supervision of AI systems by designated individuals, teams, or automated controls to ensure they operate within acceptable boundaries. Approval Workflow An approval workflow is a structured process that requires one or more human reviewers to authorise a high-risk agent action before it is executed. Human Oversight Human oversight is the ability of designated individuals to monitor, intervene in, and override an AI system's decisions or actions.
Isolation & containment
Agent Boundary An agent boundary is a hard constraint that an agent cannot cross regardless of its reasoning or the user's instructions. Agent Context Isolation Agent context isolation is the enforcement of boundaries between the execution contexts of different agents, ensuring that one agent's memory, credentials, tool state, and intermediate data cannot be read or modified by another agent without explicit permission. Agent Sandboxing Agent sandboxing is the practice of running an AI agent in an isolated environment with restricted access to systems, networks, and data. Containment Strategy A containment strategy defines how to limit the impact of an AI agent that is behaving unexpectedly, including throttling, suspending, isolating, or rolling back the agent. Containment Strategy (AI) An AI containment strategy is a set of technical and procedural controls designed to limit the potential impact of an AI agent that behaves unexpectedly or is compromised.
Kill switches & suspension
More runtime control & guardrail terms
Agent Autonomy Level Agent autonomy level describes how much independent decision-making authority an agent has, from fully supervised (every action approved) to fully autonomous (acts without human checkpoints). Agent Middleware Agent middleware is software that sits between AI agents and the systems they interact with, providing cross-cutting services such as authentication, logging, policy enforcement, retry handling, and protocol translation. Agent Scaffolding Agent scaffolding is the surrounding infrastructure and code that wraps an AI model to give it agentic capabilities, including the loop that calls tools, manages context, handles errors, and decides when to stop. Agentic AI AI systems that can autonomously plan, reason, use tools, and take actions to achieve goals with minimal human intervention. AI Change Management AI change management is the governed process of reviewing, approving, and deploying changes to AI agents, including model updates, prompt modifications, tool additions, and policy adjustments. AI Governance Board An AI governance board is a senior cross-functional committee that sets AI strategy, approves high-risk use cases, resolves escalations, and ensures the organisation's AI activities align with its values, policies, and regulatory obligations. AI Product Owner An AI product owner is responsible for the business value, scope, and requirements of an AI agent or agent-powered product. Ambient Authority Ambient authority is a security anti-pattern in which an agent inherits broad permissions from its execution environment, such as the credentials of the user account it runs under, rather than being granted only the specific permissions it needs for a task. Audit Trail An audit trail is a chronological, tamper-evident record of actions taken within a system. Autonomous Agent An autonomous agent is an AI system that can independently plan, reason, use tools, and take actions to achieve a goal with minimal or no human intervention during execution. Capability Control (AI) Capability control is the practice of limiting what an AI agent is technically able to do, through tool restrictions, output filtering, API permissions, and execution environment constraints, rather than relying solely on the agent's own safety training to prevent harmful behaviour. Code Execution Agent A code execution agent is an agent that writes and runs code in a sandboxed environment to solve analytical, computational, or data transformation tasks. Data Classification Data classification assigns sensitivity labels to data, such as public, internal, confidential, or restricted, based on its content and regulatory requirements. Evasion Attack (AI) An evasion attack crafts inputs that cause an AI agent's safety or policy filters to fail to detect policy violations, allowing harmful content, prompt injection, or unauthorised instructions to pass through governance controls. Evidence of Compliance Evidence of compliance is the documented, auditable proof that a governance control was in place and operating effectively at the time a requirement was in scope. Governance Automation Governance automation is the use of technology to enforce, monitor, and report on AI governance policies without manual intervention. Governance by Design Governance by design is the principle of embedding governance controls into AI systems from the start rather than adding them after deployment. Governance Event A governance event is a logged record of a policy evaluation, violation, escalation, or override that occurred during an agent's operation. Jailbreak (AI Agent) A jailbreak is a technique used to bypass an AI agent's safety guardrails or governance constraints, typically through crafted prompts, role-play framings, or instruction injections that cause the agent to ignore its system prompt or policy rules. Model Risk Manager A model risk manager oversees the identification, assessment, and mitigation of risks arising from the use of quantitative and AI models in business decisions. Red Teaming (Agent) Agent red teaming is the practice of having adversarial testers systematically probe an AI agent for vulnerabilities, policy bypasses, and unsafe behaviors before it reaches production. Shadow AI Shadow AI is any AI tool, model, or agent used inside an organisation without the knowledge or approval of the teams accountable for it. Tool Registry A tool registry is a centralised catalogue of the tools, APIs, and MCP servers that AI agents are approved to use within an organisation.
Questions
Common questions.
What are agent guardrails?
Runtime constraints that limit what an agent can do, what data it can access, and how it can respond, designed to hold even when the agent reasons toward an unsafe action.
When should an agent action require human approval?
When the action's consequence outweighs the latency cost of a checkpoint. High-impact, hard-to-reverse actions justify human-in-the-loop; routine ones usually do not. The entries in the human control group cover the mechanisms.