NVIDIA NCP-AAI key terms

A software system in which a language model plans, selects tools, and takes multi-step actions toward a goal with limited human step-by-step control.

The repeating cycle of observing state, reasoning about the next step, invoking a tool or action, and incorporating the result until the goal or a stop condition is reached.

A mechanism where the model emits a structured request naming an external function and its arguments, and the runtime executes it and returns the result to the model.

A machine-readable description of a callable tool, including its name, purpose, parameter types, and required fields, used by the model to form valid calls.

An architecture that separates an agent that decomposes a goal into steps from workers that carry out each step, improving reliability and allowing independent scaling of each role.

A multi-agent topology in which a coordinating agent routes subtasks to specialized agents and aggregates their outputs into a final response.

An agent whose only job is to monitor other agents, decide who acts next, resolve conflicts, and enforce completion criteria rather than perform domain work itself.

The transfer of control, context, and pending task state from one agent to another, including what history and constraints travel with it.

The common store of conversation history, intermediate results, and constraints that participating agents read and write so their work stays consistent.

Breaking a complex goal into smaller, independently executable subtasks with clear inputs, outputs, and success criteria so agents can work in parallel or in sequence.

A self-critique step in which an agent reviews its own draft or trajectory against the goal and revises before returning a result.

A separate agent or model call that scores another agent's output against explicit criteria, used to catch errors before delivery.

A predefined rule that stops an agent loop, such as goal satisfaction, a maximum step count, a timeout, or a detected repeated action.

The degree to which an agent may act without human approval, ranging from suggestion-only to fully automated execution of high-impact actions.

A control pattern requiring a person to review or approve an agent's proposed action before it is executed, typically for irreversible or high-risk operations.

A checkpoint in an agent workflow where execution pauses until a designated human or policy engine authorizes the next action.

An enforced constraint on agent inputs, outputs, or actions, such as content filters, allowlists, or schema validation, applied outside the model's own judgment.

A deterministic component that evaluates proposed agent actions against written rules and returns allow, deny, or require-approval decisions.

Granting an agent only the permissions and data access needed for its current task, and revoking them when the task ends.

Short-lived, narrowly permissioned tokens issued to an agent for a specific tool or resource instead of long-lived broad access keys.

Running agent-generated code or tool actions in an isolated environment with restricted filesystem, network, and resource access to contain failures and attacks.

An attack in which untrusted content, such as a retrieved document or web page, contains instructions that cause the agent to deviate from its intended task.

Prompt injection delivered through data the agent processes rather than through the user, for example a poisoned file, email, or web result.

Tampering with a tool's description, schema, or output so the agent is misled into unsafe calls or trusts fabricated results.

An immutable log of agent decisions, tool calls, inputs, and approvals that supports incident investigation, compliance review, and reproducibility.

Structured adversarial testing that probes an agent system for unsafe actions, data leakage, jailbreaks, and policy circumvention before deployment.

A short document describing a model's intended use, training data characteristics, evaluated limitations, and known risks, used in governance review.

A documented evaluation of how a model or agent could cause harm or error in a given deployment, with likelihood, impact, and mitigating controls.

Model Context Protocol, an open standard that lets agents discover and call external tools and data sources through a uniform client-server interface.

A pattern that fetches relevant documents at query time and supplies them to the model so answers are grounded in current, domain-specific content.

A dense numeric vector representing text meaning, produced by an encoder model and used for semantic search, clustering, and similarity comparison.

A data structure that stores embeddings and supports fast approximate nearest-neighbor search over large document collections.

Tying model output to verifiable source content, such as retrieved passages or tool results, so claims can be checked and hallucination is reduced.

An automated verification that each cited source actually supports the claim it is attached to, used as a grounding guardrail.

Constraining a model response to a defined schema, such as JSON with required fields, so downstream code can parse and validate it reliably.

NVIDIA Inference Microservices, containerized model-serving endpoints with standardized APIs, used to host models behind agent applications on NVIDIA infrastructure.

NVIDIA's framework family for building, customizing, and guarding generative AI pipelines, including training, evaluation, and safety components.

Hosting a model behind an API with batching, concurrency control, and health checks so agent requests get low-latency, reliable responses.

The fraction of GPU compute actively used during inference; low values indicate wasted capacity and high values near saturation indicate queuing risk.

Grouping multiple inference requests into one model call to raise throughput and GPU efficiency, trading some latency for capacity.

Stored attention keys and values from prior tokens that let a model avoid recomputation during generation, consuming GPU memory that limits concurrency.

The maximum number of tokens a model can attend to in one call, bounding how much history and retrieved content an agent can carry.

Persistent storage of facts, preferences, or past outcomes that an agent retrieves across sessions to avoid repeating work or losing state.

The serialized snapshot of an agent's task progress, tool results, and pending decisions, allowing pause, resume, and recovery after failure.

Periodically persisting agent state so a long-running workflow can resume from the last good point instead of restarting from scratch.

A property where repeating the same action produces the same result, essential for safely retrying tool calls after timeouts or partial failures.

Resending a failed request after increasing delays, often with jitter, to recover from transient errors without overwhelming a struggling service.

Capping request volume per client or per agent to protect model endpoints and downstream tools from overload and to keep latency predictable.

Automatically adjusting the number of model or agent worker instances based on load signals such as queue depth or GPU utilization.

A hard limit on tokens, tool calls, wall-clock time, or cost per task that prevents runaway agent loops from consuming unbounded resources.

Instrumentation that captures traces, metrics, and logs from agent runs so failures, latency, and cost can be diagnosed in production.

A recorded tree of spans showing each model call, tool invocation, and handoff within one agent run, including timing and payload metadata.

A repeatable test setup with curated tasks, expected outcomes, and scoring functions used to measure agent quality before and after changes.

Using a language model to score another model's output against a rubric, useful for open-ended tasks but requiring calibration against human labels.

A fixed set of agent test cases rerun on every prompt, model, or tool change to detect quality drops introduced by the update.