A practical glossary for AI systems, agents, coordination, authority, and evidence

A

A2A

Agent2Agent Protocol: a protocol for communication and collaboration among agent systems. Interoperability does not itself establish authorization or trust.

Agency

The practical degree to which a system may choose actions, tools, sequence, timing, or delegation. Agency should be described by actual authority, not by a conversational interface.

Agent

A software participant that can select among possible next actions from a goal, current state, evidence, policy, and feedback rather than following only one fixed prompt-response step.

Agentic AI

An AI-enabled system that observes, plans, uses bounded tools, evaluates results, updates state, and stops or escalates through a multi-step control loop.

Agentic RAG

A retrieval system in which an agent adaptively plans, reformulates, chooses sources or tools, iterates, verifies, or delegates retrieval rather than following one fixed retrieval pipeline.

AI system

The complete sociotechnical arrangement of models, data, software, people, policies, interfaces, infrastructure, and operating procedures used to produce an AI-assisted outcome.

Architecture decision record (ADR)

A durable record of an architectural decision, the alternatives considered, evidence, assumptions, consequences, owner, and conditions that should reopen the decision.

Authorization

The decision that a particular identity may perform a specific action on a specific resource under current policy and context.

Autonomy

The extent to which a system can continue without synchronous human instruction. Autonomy is task- and boundary-specific, not a single product-wide label.

B

Blackboard architecture

A coordination pattern in which participants read and write to a shared problem state or evidence space rather than communicating only through direct messages.

C

Capability

A bounded operation an actor can perform, such as reading a document, proposing a payment, or executing a deployment. Capabilities should be scoped and revocable.

Conditional release

A release decision that permits use only within explicit limits, with required remediation, monitoring, review dates, or excluded populations and actions.

Context window

The model-visible input available during one inference. It is temporary working context, not a durable or authoritative memory system.

Control loop

The sequence by which a system observes, interprets, plans, authorizes, acts, verifies, remembers, and stops or continues.

Coordination cost

The added latency, tokens, messages, state, conflict resolution, evaluation, and operational burden created when work is distributed among multiple participants.

D

Delegation

A principal granting an agent or subordinate actor a narrower authority to act for a defined purpose, period, resource set, and consequence limit.

Deterministic workflow

A process whose steps and transitions are expressed in code, rules, or a state machine rather than chosen probabilistically at runtime.

E

Effect verification

Checking the real downstream result of an action instead of treating a successful API response or model statement as proof that the intended business state exists.

Episode

One bounded execution or case, including its trigger, evidence, decisions, actions, state transitions, reviews, and outcome.

Evaluation dataset

A versioned collection of representative, edge, adverse, and prohibited cases with expected behavior or adjudication criteria used to test an AI system.

Evaluation harness

The repeatable software and procedure that runs cases, captures traces, scores outcomes and trajectories, and produces reviewable evidence.

G

Goal hijacking

A failure in which untrusted content, another actor, or a conflicting instruction changes the agent's effective objective away from the authorized goal.

Groundedness

The degree to which an output is supported by the permitted source evidence. Groundedness does not by itself establish completeness or correctness.

H

Human authority

The named decision rights that remain with accountable people, including approval, override, refusal, appeal, escalation, and acceptance of residual risk.

Human in the loop

A configuration in which a person participates during execution, such as by supplying judgment, correcting state, or approving a consequential step.

Human on the loop

A configuration in which a person supervises operations and can intervene, pause, or revoke authority without approving every ordinary step.

I

Idempotency

The property that repeating an operation with the same idempotency key does not create an additional unintended effect.

Identity

The verifiable representation of the human, workload, agent, service, or organization responsible for a request or action.

Indirect prompt injection

An instruction embedded in external content that attempts to redirect model or agent behavior when the content is retrieved or processed.

M

Machine intelligence

An editorial umbrella for systems that sense, interpret, predict, generate, plan, coordinate, or act with machine assistance, including the surrounding authority and evidence architecture.

MCP

Model Context Protocol: a protocol for connecting AI applications to context, resources, prompts, and tools. Protocol compatibility does not grant business authority.

Memory

Durable information made available across steps or runs. Memory requires provenance, scope, retention, supersession, access control, and separation from authoritative policy.

Multi-agent system

A system in which multiple software agents or participants coordinate, specialize, challenge, negotiate, or allocate work to produce a joint result.

O

Observability

The ability to understand system behavior from traces, metrics, events, state, and evidence sufficient to investigate outcomes, cost, failures, and side effects.

Orchestration

The mechanism that routes work, sequences steps, manages state, handles retries, invokes tools, and determines completion or escalation.

P

Policy decision point

The component that evaluates identity, action, resource, context, and rules to decide whether a requested operation is permitted.

Provenance

Information about the origin, authority, observation time, transformation, version, and custody of evidence, memory, or an output.

Q

Quorum

A threshold of participants or signals required before a distributed system changes state. Quorum is a coordination mechanism, not automatic accountability.

R

RAG

Retrieval-augmented generation: a system that retrieves source material and provides it to a generative model to support an answer or artifact.

Release gate

A documented decision point that compares evidence against explicit criteria and records release, conditional release, hold, or stop.

Residual risk

Known risk that remains after controls and is explicitly accepted, transferred, avoided, or scheduled for remediation by an authorized owner.

Rollback

A tested path for returning code, configuration, model, prompt, data, or authority to a known acceptable state after an unsafe change.

S

State

The authoritative representation of workflow progress and business facts needed to resume, reconcile, or review an execution.

Stigmergy

Indirect coordination through changes to a shared environment, where one participant's trace or artifact influences later participants without direct messaging.

Supervisor agent

A coordinating agent that assigns, routes, or reviews work performed by other agents. A supervisor remains another probabilistic component and should not be treated as policy authority.

Swarm intelligence

Decentralized coordination in which many relatively simple participants follow local rules and feedback, producing system-level behavior without one central planner.

T

Tool

A bounded external capability exposed to a model or agent, such as search, database access, messaging, code execution, or transaction proposal.

Trace

A correlated technical record of the operations and relationships in an execution, ideally connected to the business case and human decision record.

Trajectory

The ordered path of observations, plans, messages, tool calls, state transitions, reviews, and stop conditions taken during one execution.

U

Unknown state

A condition in which the system cannot determine whether an attempted external effect succeeded, failed, or partially completed and must reconcile before retrying.

No matching terms. Clear the filter to show the complete glossary.