AI Agent Glossary

A communication protocol enabling direct agent-to-agent interaction, message passing, and coordination without human mediation.

An autonomous software entity capable of perceiving its environment, reasoning about information, making decisions, and taking actions to achieve specific goals.

Systematic testing and measurement of agent performance against benchmarks, success criteria, and quality metrics.

AI systems designed to act autonomously with agency—the ability to make decisions, take action, and adapt behavior based on goals and environmental feedback.

A framework enabling conversations between multiple LLM agents and human users for collaborative problem-solving.

An agent capable of independent operation with minimal human intervention, making decisions and executing actions autonomously.

AWS managed service for running agents that automatically handle orchestration, memory, and integration with AWS services.

Storing frequently accessed results temporarily to avoid redundant computation and improve response times.

A prompting technique where an AI is instructed to break down complex problems into intermediate reasoning steps, improving solution quality.

Anthropic\

An approach training AI systems to follow explicit principles and guidelines, improving safety and alignment with human values.

Strategies for maintaining, updating, and optimizing the information available to agents within context limits.

The maximum amount of text (in tokens) that a language model can process at once, limiting how much history and context it can access.

A framework specifically designed for building multi-agent systems with role-based agents, tasks, and hierarchical management.

Tracking requests and operations across multiple services and components, enabling visibility into complex distributed workflows.

Numerical vector representations of text that capture semantic meaning, enabling comparison and retrieval based on semantic similarity.

Memory of specific events, experiences, and interactions, typically time-stamped and contextual, useful for learning from experience.

Mechanisms for detecting, managing, and recovering from failures gracefully without system collapse.

Providing a small number of examples in a prompt to teach LLMs new patterns or behaviors without fine-tuning.

Training an LLM on task-specific data to specialize its behavior, improving performance on particular domains or tasks.

A mechanism where an LLM generates structured outputs (typically JSON) that describe which function to call with what parameters, enabling programmatic tool invocation.

Breaking down high-level goals into smaller, manageable subgoals that are easier to achieve and reason about.

Google\

An agent architecture where nodes represent agents/tasks and edges represent dependencies, enabling complex workflow management.

Safety mechanisms and constraints that limit agent behavior, preventing harmful outputs and ensuring compliance with guidelines.

When LLMs generate false, incorrect, or fabricated information presented as fact, a critical safety concern.

A multi-agent pattern where agents explicitly pass work to each other based on specialization, creating a workflow of transfers.

A multi-level agent structure where higher-level agents coordinate lower-level agents, creating nested decision-making hierarchies.

A control pattern where significant agent decisions require human review and approval before execution.

A standard for describing and validating JSON data structures, specifying properties, types, constraints, and requirements.

A structured representation of entities, relationships, and facts organized as a graph, enabling complex reasoning and relationship queries.

A popular framework for building applications with large language models, providing abstractions for chains, memory, and external integrations.

An extension of LangChain enabling graph-based workflow definition with cycles, state management, and complex agent patterns.

The time delay between when an action is initiated and when results are returned, critical for user experience.

Classification framework describing agent autonomy from fully manual to fully autonomous decision-making.

A data indexing framework for building RAG systems, providing tools for data ingestion, indexing, and retrieval optimization.

Persistent storage for important information, patterns, and knowledge accumulated over time that persists beyond individual interactions.

An open protocol for standardizing how AI agents connect to data sources, tools, and other services, enabling interoperability.

Architectures where multiple independent agents work together, each with specialized capabilities, to solve problems collaboratively.

A workflow automation platform enabling visual agent and workflow creation with native LLM integration and 400+ pre-built nodes.

The capability to understand and debug complex systems through comprehensive logging, metrics, and visibility into system behavior.

An open standard for describing REST APIs in a machine-readable format, enabling automated documentation, client generation, and tool integration.

The coordination and management of multiple agents or components working together to accomplish complex workflows and achieve shared objectives.

An execution pattern where multiple agents or tasks run concurrently, improving performance by utilizing multiple resources simultaneously.

The process by which an agent breaks down complex goals into smaller subgoals and sequences of actions needed to achieve them.

The practice of crafting inputs to LLMs to achieve desired outputs, including instruction clarity, examples, and formatting.

An attack where malicious input is crafted to manipulate LLM behavior, potentially bypassing safety measures and instructions.

A technique combining information retrieval from external sources with generation, allowing agents to access relevant knowledge without fine-tuning.

Controlling the frequency of requests to prevent overload, manage costs, and ensure fair resource allocation.

Reasoning + Acting framework that combines explicit reasoning steps with action steps, enabling agents to interleave thinking with tool usage.

The cognitive process by which an agent analyzes information, draws logical conclusions, and generates justifications for decisions and actions.

Adversarial testing approach where teams actively try to break agent security and identify vulnerabilities before deployment.

An agent capability to evaluate its own outputs, identify errors or inconsistencies, and self-correct before returning final results.

Techniques for ensuring LLM outputs conform to desired formats (JSON, XML, markdown, structured data).

A training technique using human feedback to fine-tune model behavior, improving alignment with human preferences.

Isolating agent execution environments to limit the impact of potential errors or malicious code, containing damage.

A lightweight orchestration framework integrating LLMs with plugins, enabling flexible skill composition and function integration.

Search technique that finds results based on semantic meaning rather than keyword matching, using embeddings.

An execution pattern where agents or tasks execute one after another in a defined order, with output from one feeding into the next.

Temporary working memory that holds recent context, current task information, and intermediate results during active processing.

A lightweight, Hugging Face-based framework for building agents with minimal dependencies, focusing on simplicity and transparency.

An architecture where agents operate as state machines with defined states, transitions, and actions triggered by events or conditions.

Sending results incrementally as they\

An agent architecture where a central supervisor agent delegates work to specialized worker agents and coordinates their activities.

An agent architecture inspired by natural swarms where simple agents following local rules create emergent global behaviors.

Initial instructions provided to an LLM that define its role, behavior, guidelines, and constraints for the entire conversation.

The process by which an agent breaks down goals into concrete actions and executes them systematically to completion.

A setting controlling randomness in LLM outputs; lower values (0-0.3) produce deterministic outputs, higher (0.7-1.0) increase creativity.

The allocated number of tokens an agent can consume per time period, controlling costs and managing resource usage.

The fundamental units that LLMs process, typically representing subword units that are smaller than words but larger than characters.

The process of connecting agents to external tools, APIs, and services enabling extended capabilities.

Formal definitions of available tools specifying parameters, return types, descriptions, and usage patterns in structured formats.

The capability of an AI agent to identify, select, and utilize external tools (APIs, functions, code executors) to accomplish tasks beyond its native knowledge.

A sampling parameter controlling diversity by including only tokens with cumulative probability up to p, balancing quality and variety.

A database designed to store and retrieve high-dimensional vector embeddings efficiently, enabling semantic search and similarity matching.

The active memory used during task execution to hold intermediate values, processing state, and temporary information needed for immediate operations.

Using an LLM to perform tasks without providing specific examples, relying on the model\