Machine intelligence, agentic AI, and AI agent glossary

Reviewed terminology

Agent evaluation

The systematic testing of an agent or multi-agent system across representative scenarios, including task quality, plan adherence, tool use, coordination, safety, latency, cost, and recovery.

#

Agent incident review

An evidence-linked reconstruction of an agent failure covering trigger, state, messages, tools, policy decisions, human actions, side effects, containment, recovery, and corrective action.

#

Agent memory

The governed storage and retrieval of task state, prior events, durable facts, and reusable context used by one or more agents across steps or sessions.

#

Agent observability

The ability to reconstruct agent state, messages, plans, model calls, tool calls, policy decisions, approvals, costs, errors, and outcomes from durable telemetry.

#

Agent registry

An inventory of deployed or approved agents, including purpose, owner, model, tools, data access, permissions, dependencies, risk tier, evaluation status, and lifecycle state.

#

Agent swarm maturity model

A staged assessment of how consistently a multi-agent system implements architecture, authority, evaluation, operations, memory, and evidence.

#

Agentic AI

AI designed to pursue a defined objective through multiple steps, select or invoke tools, maintain task state, and adapt its next action within explicit limits.

#

Agentic control plane

The policy and operations layer that governs agent identity, permissions, tool access, model routing, budgets, state, approvals, telemetry, evaluation, and shutdown independently of the agents doing the work.

#

Agentic readiness

The evidenced condition in which a bounded agentic workflow has sufficient business definition, architecture, authority, evaluation, operations, and recovery to support its next decision.

#

Agentic swarm

A coordinated multi-agent system whose specialized agents can divide, delegate, review, and recombine work under shared policy and control.

#

Agentic swarm management

The operating discipline for designing roles, assigning work, controlling permissions and budgets, managing shared context, evaluating behavior, handling failures, and preserving human authority across an agent swarm.

#

Agentic system card

A version-linked summary of an agentic system’s purpose, owners, scope, architecture, dependencies, authority, evaluation evidence, operations, limitations, incidents, and change history.

#

Agentic threat taxonomy

A shared classification of attack, misuse, failure, identity, tool, memory, communication, human-trust, cascade, and rogue-agent risk patterns used to structure threat modeling and tests.

#

AgentOps

The operating discipline for agentic systems, covering goals, plans, state, tools, identities, permissions, handoffs, human authority, side effects, evaluation, budgets, incidents, and recovery.

#

AI agent

A bounded software actor that receives an objective, uses a model and available context to choose actions, invokes approved tools, and returns evidence about what it did.

#

AI bill of materials (AIBOM)

A structured record of the models, prompts, datasets, tools, libraries, services, policies, and dependencies used by an AI system.

#

AI governance evidence register

A traceability record linking a selected requirement, objective, risk, or policy statement to scope, owner, control, evidence artifact, verification result, exception, approval, and next review.

#

AI management system (AIMS)

An organization-wide set of interrelated policies, objectives, processes, roles, controls, monitoring, review, and improvement activities for the responsible development, provision, or use of AI systems.

#

AI risk management framework

A structured set of concepts and outcomes used to identify, govern, map, measure, manage, communicate, and revisit AI risks across a lifecycle and organizational context.

#

AI scale gate

An evidence-backed decision process that determines whether an AI-enabled business process should proceed, remediate, rebid, replace, narrow, or stop before broader investment.

#

Architecture decision record (ADR)

A durable record of an architecture context, considered alternatives, decision, trade-offs, evidence, ownership, and conditions that require reconsideration.

#

Artificial intelligence (AI)

The broad field of computational systems that perform tasks associated with perception, prediction, language, reasoning, decision support, or action.

#

Cascading agent failure

A local error or unsupported state that crosses a boundary and is amplified through downstream agents, tools, memory, or decisions.

#

Cascading failure

A failure in which one agent’s incorrect state or output is accepted by downstream agents and propagates into additional decisions or actions.

#

Claims registry

A structured record that classifies public statements by evidence status, permitted use, source basis, validation needs, and prohibited forms.

#

Context engineering

The design of what information an agent receives, when it receives it, how it is selected or compressed, and how unrelated or sensitive context is isolated.

#

Context rot

A loss of retrieval and reasoning quality that can occur when an agent is given excessive, stale, conflicting, or weakly prioritized context.

#

Corrections register

A versioned record that preserves earlier wording, states the current corrected or clarified language, identifies affected content, and cites the supporting source.

#

Evaluation gate

A deterministic checkpoint that blocks promotion or execution until evidence meets specified thresholds for quality, safety, policy, cost, and recoverability.

#

Execution contract

A structured request for an agent or tool action that states the objective, inputs, authority, policy, evidence, expected side effects, timeout, and failure behavior.

#

Execution sandbox

An isolated environment that limits process, filesystem, network, credential, and production access for agent-generated code or high-risk tool actions.

#

Governed autonomy

A condition in which machine intelligence can act without step-by-step human direction only inside explicit, observable, reversible, and enforceable boundaries.

#

Human authority boundary

A defined threshold where machine discretion ends and a named person or accountable role must review, authorize, modify, or stop a consequential action.

#

Independent AI Scale Gate

A buyer-side decision engagement that evaluates process economics, technical performance, implementation readiness, governance evidence, and alternatives before recommending scale, remediation, replacement, rebid, or stop.

#

Least agency

The principle that an agent should receive only the discretion, tools, data, duration, and action scope required for its current task.

#

LLMOps

The engineering and operating discipline for developing, deploying, evaluating, observing, and governing applications that use large language models.

#

Machine intelligence

A complete operational system in which models, deterministic software, data, tools, memory, policies, evaluation, and human authority work together to perform bounded tasks.

#

Model Context Protocol (MCP)

An open protocol for connecting AI applications to tools and context sources. Using it safely still requires authentication, authorization, input controls, isolation, logging, and tool-specific policy.

#

Multi-agent system

A system in which two or more specialized agents coordinate through defined roles, messages, shared state, handoffs, and completion rules.

#

Non-human identity

A machine, workload, service, bot, or agent identity that accesses systems without being a person and therefore requires explicit ownership, credentials, scope, and lifecycle controls.

#

Official source ledger

A reviewed registry of primary sources with publisher, version or status, publication and review dates, summary, scope, and public-use boundary.

#

Production readiness

The demonstrated ability of an AI system to meet defined operational, security, governance, quality, cost, ownership, and recovery requirements under representative conditions.

#

Protocol boundary

The explicit line between behavior standardized by an interoperability protocol and the identity, policy, evaluation, authority, recovery, and operating responsibilities implemented around it.

#

Semantic boundary

A handoff point where intent, meaning, confidence, state, or completion must be interpreted correctly by another agent, tool, or person.

#

Semantic circuit breaker

A control that pauses or stops a workflow when meaning, confidence, policy, or expected state diverges—even when the underlying software call technically succeeds.

#

Shadow agent

An agentic workflow deployed without recognized ownership, security review, policy enforcement, or inclusion in the organization’s system inventory.

#

Source provenance

The traceable origin, transformation, review status, and evidence boundary of information used in a page, answer, decision, or model context.

#

State-machine orchestration

An orchestration pattern in which workflow states and allowed transitions are explicitly represented, tested, checkpointed, and observed.

#

Telemetry semantic conventions

Shared names and meanings for trace, metric, event, and log fields that make observability data more portable and comparable across instrumentation and backends.

#

Vendor neutrality

A decision process whose evidence, compensation, architecture, and exit design allow credible comparison among vendors, models, custom solutions, conventional automation, and no implementation.

#

Workflow automation

Deterministic software that moves work through explicit rules, states, integrations, and exception paths.

#