AI-500 key terms

A software component that uses a language model to reason about a goal, decide on actions, call tools, and produce results with some autonomy rather than only answering a single prompt.

A coordinating agent that decomposes a request, decides which specialist agents or tools to invoke, passes context between them, and assembles the final response for the user.

A narrower agent invoked by a parent or orchestrator agent to handle one bounded task, usually with a smaller tool set and its own instructions and context.

A recurring structural design for splitting work among several agents, such as supervisor, sequential pipeline, concurrent fan-out, or debate, chosen to match task dependencies and failure tolerance.

Transferring control of a conversation or task from one agent to another, along with the context the receiving agent needs to continue without asking the user to repeat information.

The declared responsibility, scope, and boundaries assigned to an agent, which shapes its instructions, allowed tools, and what it is permitted to decide or escalate.

The system-level prompt text that defines an agent's persona, goals, constraints, output format, and escalation rules, and that is applied on every turn of its reasoning.

The mechanism by which a model emits a structured request for an external function, such as an API or database lookup, whose result is returned to the model for further reasoning.

The machine-readable description of a callable tool, covering its name, purpose, parameter types, and required fields, that the model uses to decide when and how to call it.

A common store of conversation history, intermediate results, and task status that multiple agents read and update so their work stays consistent across handoffs.

An ordered record of messages and tool interactions belonging to one session, used to preserve context and to let agents resume or audit a multi-turn exchange.

The maximum amount of text, in tokens, that a model can consider at once; managing it is essential because multi-agent transcripts grow quickly and can exceed the limit.

Deliberately selecting, summarizing, and ordering what each agent sees, so it receives relevant state and instructions without flooding the model with unnecessary history.

A planned cap on tokens consumed per request, agent, or workflow, used to control cost and latency and to force summarization before limits are reached.

Persisting an agent workflow's progress at defined points so a long-running or failed run can resume from the last good state instead of restarting from the beginning.

A property of an operation that produces the same result when repeated, which lets agents safely retry tool calls after timeouts without duplicating side effects like orders or payments.

Rules governing automatic re-attempts of failed calls, including which errors are retryable, how many attempts are allowed, and how long to wait between them.

A protective pattern that stops sending requests to a failing dependency after repeated errors, failing fast until a recovery period passes, to prevent cascading agent failures.

A coordination style in which several agents exchange messages in one shared conversation, with a manager or rule deciding who speaks next and when to stop.

A software library or service that supplies the plumbing for agents, including model clients, tool registration, message passing, state management, and orchestration primitives.

A provisioned hosting endpoint for a language model in a cloud environment, with a chosen model version, capacity, and region that agents call at runtime.

Supplying an agent with authoritative external data, such as documents or search results, so its answers rest on retrieved facts rather than only on model training.

A technique that retrieves relevant content from a knowledge source and inserts it into the model's context before generation, improving factual accuracy and traceability.

A data structure storing embedded representations of content so that semantically similar items can be found quickly during retrieval for grounding an agent's response.

A confident model output that is fabricated or unsupported by available evidence, which multi-agent designs mitigate through grounding, verification agents, and explicit refusal instructions.

An agent whose job is to judge another agent's output against criteria such as accuracy, completeness, or policy compliance, producing a score or corrective feedback.

Using a language model to score or compare outputs against a rubric, which scales qualitative evaluation but requires calibration against human-labeled examples.

A curated set of representative inputs with expected outputs or grading criteria, used to measure agent quality consistently and to compare changes over time.

A human-approved reference answer for a test case that serves as the benchmark against which an agent's output is compared during evaluation.

Re-running a fixed evaluation suite after changes to prompts, models, or tools to detect quality drops before the updated solution reaches production.

Ongoing automated scoring of live or sampled production interactions so quality drift, new failure modes, and unsafe outputs are detected between formal release cycles.

The ability to understand an agent system's internal behavior from its outputs, achieved through traces, logs, and metrics that show each step of a multi-agent run.

Recording a single request's path across agents, tools, and services as linked spans, so latency and errors can be attributed to a specific step.

The metrics, logs, and events emitted by a running system and collected centrally, used to monitor health, cost, and behavior of agents in production.

The count of input and output tokens consumed by model calls, tracked per agent and per request because it drives both cost and latency in multi-agent workflows.

The total model and tool spend divided by successfully completed tasks, a business-level metric for judging whether a multi-agent design is economically viable.

The authenticated identity an agent uses to access resources, ideally distinct from a human user's credentials, so permissions can be scoped and audited per agent.

A platform-issued identity for a hosted workload that obtains tokens automatically, removing the need to store secrets in code or configuration for agent-to-service calls.

Granting each agent and identity only the permissions its task requires, limiting the blast radius if a prompt injection or bug causes unintended tool calls.

A control that constrains agent behavior, such as input filtering, output validation, allowed-tool lists, or human approval, to keep actions within policy and safety bounds.

A service-side classifier that screens model inputs and outputs for harmful categories and can block or annotate them before they reach users or downstream agents.

An attack that hides malicious instructions in content an agent processes, attempting to override its instructions or trigger unauthorized tool calls and data exposure.

A design in which a person reviews or approves an agent's proposed action before it executes, used for high-impact operations that should not run fully autonomously.

Storing and rotating credentials such as API keys in a protected vault rather than in code or prompts, so agents can authenticate without exposing sensitive values.

A reserved model capacity model that guarantees a fixed processing rate for predictable latency and cost, in contrast to pay-per-token consumption-based deployment.

A structured assessment of an agent solution for fairness, safety, transparency, and privacy risks before deployment, with mitigations documented and tracked.