Loading…
Choose a reading style and confirm your keyword above, then open or browse below. Carousel is the default.
Default: Carousel — switch below anytime.
View as
No keyword yet — type a term above or use Look up a term below.
Opens in your selected reading style above.
Search the full glossary · opens in your selected style above.
Tab or tap the match line to accept · tap your selection above or Open in … here.
A software-based entity that perceives its environment, reasons over context, and takes actions to achieve specific objectives. Unlike static models or simple chatbots, agents can use tools, maintain state, and execute multi-step plans autonomously.
AI systems designed to autonomously plan, decide, and act toward goals rather than simply respond to prompts. Agentic AI combines LLMs, tools, and knowledge to enable end-to-end workflows with a sense of purpose and context.
A task execution pattern where an LLM operates with autonomy, making decisions about which actions to take and in what order. Contrasts with deterministic pipelines where every step is predefined.
An agent capable of operating independently over extended periods without human intervention. Can set subgoals, recover from errors, and adapt to unexpected situations.
A neural network trained on vast text corpora that understands and generates human language. LLMs serve as the "brain" of most modern AI agents. Examples: GPT-4, Claude, Llama, Gemini, DeepSeek, Qwen.
A general-purpose model that can be adapted to many downstream tasks. Foundation models enable flexibility across use cases without task-specific training.
The maximum amount of text (measured in tokens) an LLM can process in a single request. Ranges from 4K to 200K+ tokens. Constrains how much history, documents, and instructions an agent can include.
The fundamental unit of text processed by LLMs. A token is roughly ¾ of a word. Token count determines context window usage and API costs.
An agent's ability to work toward defined outcomes rather than isolated tasks. Enables intelligent sequencing of actions across systems.
A model specifically optimized for multi-step reasoning, planning, and problem-solving rather than simple text generation. Reasoning models use extended chain-of-thought before producing answers. Examples: o1, o3, o4-mini (OpenAI), Gemini 2.5 Pro (Google), Claude 3.7 Sonnet with extended thinking (Anthropic), DeepSeek-R1.
An AI-powered software development workflow where an LLM acts as an autonomous coding agent — reading codebases, planning changes, writing code, running tests, and iterating on failures. Tools: Cursor, GitHub Copilot, Windsurf, Aider, Cline. Represents a shift from autocomplete to autonomous software engineering.
An AI assistant that works alongside humans in a shared environment, offering suggestions, completions, and recommendations while keeping the human in control. Distinguished from fully autonomous agents by the expectation of human review.
See MCP in section 7. An open standard for connecting AI agents to external tools, data sources, and services via a unified protocol. ---
A prompting technique where the model shows step-by-step reasoning before arriving at a final answer. Improves performance on complex tasks by making intermediate steps explicit.
An agent architecture that interleaves reasoning traces with action execution. Pattern: Thought → Action → Observation → Thought → Action → … Makes agent behavior interpretable and effective.
Breaking down complex goals into sequences of achievable subgoals and actions. Can be done upfront (complete plan) or interleaved with execution (dynamic replanning).
Extension of chain-of-thought where the agent explores multiple reasoning paths simultaneously, evaluating and pruning branches to find optimal solutions.
Generating multiple reasoning paths and selecting the most frequent answer. Reduces dependence on any single reasoning chain for improved reliability.
Agents analyze their own outputs, reasoning, or actions to identify errors and improve. Enables self-correction without external feedback.
The logic layer that evaluates context, constraints, and options to determine next actions. Strong reasoning is foundational to reliable agent behavior.
The component responsible for breaking down high-level goals into executable steps. Enables agents to handle complex, multi-stage workflows.
The component that carries out planned actions by invoking tools, APIs, or workflows. Translates reasoning into real-world impact.
An internal representation agents use to simulate and predict outcomes, enabling effective planning.
An agent's internal estimate of the current situation, especially when complete information is unavailable.
A technique where the model allocates additional tokens to internal reasoning before producing a final answer. Enables deeper analysis on complex problems at the cost of latency and token usage. Used in o1/o3, Claude 3.7 Sonnet, and Gemini 2.5 Pro.
Allocating more computational resources during inference (at query time) rather than training time. By spending more "thinking budget" on hard problems, models can achieve better accuracy without larger training runs.
A model trained to predict human preferences or quality scores for model outputs. Used in RLHF (Reinforcement Learning from Human Feedback) to align language models with desired behavior.
A method (pioneered by Anthropic) where AI models are trained to self-critique and revise their outputs based on a set of principles (a "constitution"), reducing reliance on human feedback for alignment.
Generating chain-of-thought reasoning traces using a model itself, then using those traces to train or fine-tune smaller models. Enables distillation of reasoning capabilities.
The core execution cycle of an agent: observe → reason → act → observe → ... Continues until the goal is achieved or a stopping condition is met. May include reflection, replanning, and tool use within each iteration.
The theoretical capability of an AI system to improve its own code, prompts, or reasoning strategies over time. A key aspiration in agentic AI research. ---
The mechanism by which agents invoke external functions or APIs. The LLM generates structured output specifying which tool to call and with what parameters. Bridges text-based reasoning with real-world actions.
A discrete capability made available to an agent, defined with a name, description, and parameter schema. Examples: web search, code execution, database queries, API calls.
The complete set of actions available to an agent, including all tools and their possible parameters. Larger spaces provide more capability but increase selection difficulty.
The step an agent takes to influence its environment — sending an email, executing code, calling an API.
Model responses in a specific format (JSON, XML) that can be reliably parsed by code. Enables programmatic processing and tool calling.
Connecting agent outputs to verifiable external sources. Grounded agents cite sources, retrieve facts, and verify claims. Reduces hallucination.
The result returned to an agent after executing an action. Informs the next reasoning step; may include data, text, errors, or state changes.
Connecting agents to enterprise systems (ticketing, billing, data platforms). Unlocks automation at scale.
An agent that interacts with software by seeing screenshots, clicking buttons, typing text, and navigating graphical interfaces — mimicking human computer interaction. Examples: Claude Computer Use, OpenAI Operator, Google Mariner.
Agents that control web browsers to navigate pages, fill forms, extract data, and perform multi-step web workflows. Tools: Playwright, Puppeteer, Stagehand, Browser Use.
An agent tool that executes code (typically Python) in a sandboxed environment. Enables data analysis, visualization, file processing, and complex computation as part of agent workflows.
Extending grounding to require agents to cite specific sources, page numbers, or timestamps for claims. Enables users to verify information independently.
Running agent-generated code or tool calls in isolated environments to prevent unintended side effects. Critical safety mechanism for autonomous agents. ---
Storage mechanisms that allow agents to retain and recall information over time. Enables continuity, personalization, and learning.
Information held in the current context window — recent conversation turns, active task state, retrieved documents. Limited by context window size.
Persistent storage of facts, preferences, and historical interactions surviving beyond individual conversations. Typically uses vector databases for semantic retrieval.
Storage of complete experiences (situation, actions, outcomes). Enables learning from past successes and failures by recalling similar situations.
A database optimized for storing and retrieving high-dimensional vectors (embeddings). Powers semantic search. Examples: Pinecone, Weaviate, Chroma, Qdrant, Milvus.
A numerical representation of text as a vector in high-dimensional space. Semantically similar content produces similar embeddings. Models: OpenAI text-embedding-3, sentence-transformers.
Search based on meaning rather than exact keyword matching. Uses embeddings to find conceptually similar content.
Dividing documents into smaller segments for embedding and retrieval. Chunk size trades off retrieval precision (smaller) vs context (larger).
A second-stage retrieval process that reorders initial search results using a more sophisticated model. Improves retrieval quality.
Saving agent state at specific points during execution. Enables resumption after interruption, debugging, and human review.
Tracking agent context and progress across interactions. Ensures continuity.
Handling discrete user-agent interactions securely and efficiently.
Periodically compressing and organizing accumulated memories to retain only the most important information. Mimics human memory consolidation during sleep.
Persistent storage of tool schemas, usage patterns, and results that agents can reference across sessions. Reduces repeated discovery and improves efficiency. ---
A pattern that enhances LLM responses by retrieving relevant documents before generation. Reduces hallucination and provides access to current information. Flow: Query → Retrieve → Augment prompt → Generate.
RAG in an agentic context where the agent retrieves data iteratively to refine answers across multiple steps.
A curated collection of enterprise content (FAQs, policies, documentation) that agents query for grounded responses.
RAG enhanced with knowledge graph relationships. Instead of flat document retrieval, Graph RAG traverses entity-relationship graphs to find contextually connected information. Useful for multi-hop reasoning questions.
Combining keyword-based (BM25) and semantic (vector) search for more robust retrieval. Handles both exact-match queries and conceptual searches.
An agent deciding how to handle a user query: direct LLM response, RAG retrieval, tool call, or escalation. Improves latency and relevance by selecting the optimal path.
A technique where the model learns to decide when and what to retrieve during generation, rather than always retrieving. Uses special tokens to trigger retrieval when the model lacks knowledge.
Retrieval where the agent iteratively refines queries, evaluates results, and decides whether to retrieve more information. Goes beyond single-pass RAG for complex research tasks. ---
An architecture where multiple specialized agents collaborate to solve complex problems. Agents may have different roles, tools, or expertise.
The coordination layer managing multiple agents — routing tasks, aggregating results, handling dependencies. Can be centralized (manager agent) or decentralized (agents negotiate directly).
An agent that delegates tasks to worker agents, monitors progress, and synthesizes results. Enables hierarchical decomposition of complex tasks.
Many simple agents with limited individual capability produce sophisticated collective behavior through local interactions. Inspired by biological systems (ant colonies).
Unexpected complex patterns that arise when multiple agents interact under simple rules.
The ability to run multiple agent processes simultaneously. Required for high-volume environments.
Coordinating multiple agents, tools, and workflows into a cohesive system end-to-end.
Conventions by which agents exchange information, make requests, and share results. May use natural language, structured messages, or specialized formats. ---
Open standard (now under Linux Foundation's Agentic AI Foundation) for AI applications/agents to access external tools, resources, and prompts. JSON-RPC based. Three actors: Host, MCP Client, MCP Server. Solves tool and context integration.
Google's open standard for how agents communicate with each other. Uses Agent Cards for discovery, tasks with lifecycle states, and supports streaming/async communication. Solves inter-agent delegation.
Another agent communication standard (one version deprecated and absorbed into A2A). Focused on structured messaging between agents.
An alternative protocol for agent networking and discovery.
Swarms framework protocol for deploying and managing agents as distributed services.
A JSON document (A2A) describing an agent's ID, supported modalities, tool capabilities, authentication schemes, and endpoints. Published at `/.well-known/agent-card.json`.
The fundamental unit of work in A2A. Has lifecycle states: submitted, queued, running, succeeded, failed. Agents communicate via task delegation and status updates.
Process where agents or hosts discover what tools, resources, and features are available before interaction.
A centralized registry or platform where agents can discover, evaluate, and connect with other agents or services. Enables a "directory" model for agent interoperability.
The ability for agents using different communication protocols (MCP, A2A, etc.) to work together through adapters or translation layers.
The sequence of task hand-offs between agents in a multi-agent system. Each delegation includes context, expectations, and success criteria. ---
Instructions establishing an agent's role, capabilities, constraints, and behavior guidelines. Persists across conversation turns.
The craft of designing prompts that elicit desired behavior from language models. Encompasses instruction clarity, example selection, and constraint communication.
Including examples of desired input-output pairs in the prompt to guide behavior. Enables rapid task adaptation.
Asking the model to perform a task without providing examples, relying solely on instructions.
Asking the model to explain its reasoning step-by-step before arriving at an answer. Improves accuracy on math, logic, and multi-step problems.
Extending CoT by exploring multiple reasoning paths simultaneously, evaluating branches, and pruning suboptimal paths. Used for complex planning tasks.
Fine-tuning a model on structured instruction-response pairs to improve its ability to follow complex, multi-step instructions.
Breaking a complex task into sequential prompts where each step's output feeds into the next. Enables decomposition of hard problems into manageable subtasks.
Crafting prompts that give agents autonomy, context, and tool access while maintaining guardrails. Includes system prompts, tool schemas, and behavioral constraints. ---
Widely adopted framework for building LLM-powered agents. Provides modular components for chaining prompts, integrating APIs, managing memory, and designing multi-step workflows.
Extension of LangChain for building stateful, multi-actor agent applications with graph-based orchestration.
Microsoft's open-source framework enabling LLM-based agents to collaborate conversationally. Agents delegate tasks and refine results through dialogue.
Framework for multi-agent collaboration through "crews" — specialized agent groups with role assignment and orchestration. Popular for domain-specific AI teams.
Hugging Face's minimalist framework for quickly building and experimenting with AI agents. Prioritizes simplicity and rapid prototyping.
Design where agents respond to events or triggers rather than static requests. Supports proactive automation.
End-to-end execution of business processes without manual intervention. Agentic workflows go beyond rule-based automation. ---
Constraints and checks preventing agents from taking harmful, unauthorized, or undesirable actions. Include content filters, action allowlists, rate limits, and approval requirements.
Architecture where humans review and approve certain agent decisions before execution. Provides safety at the cost of latency.
Humans supervise agents and intervene only when thresholds are crossed. Enables scale with oversight.
Adversarial testing where teams attempt to make AI systems produce harmful outputs or take dangerous actions. Identifies vulnerabilities before deployment.
When a model generates plausible-sounding but factually incorrect information. Persistent challenge requiring RAG, grounding, and verification.
Ensuring AI systems behave according to human values and intentions. Covers technical safety and ethical considerations.
Manipulating inputs to trick AI models into producing harmful or unintended outputs. Can bypass security filters or leak data.
Feeding malicious data into training sets so models learn wrong behaviors. Degrades performance or creates vulnerabilities.
Circumventing an AI model's safety restrictions through crafted prompts to make it produce disallowed content.
Policies, controls, and processes ensuring responsible agent behavior. Critical for enterprise adoption.
Adherence to regulatory, legal, and organizational standards. Requirements vary by industry.
The ability to understand and justify why an agent took a particular action. Critical for compliance, trust, and debugging.
Visibility into agent behavior, decisions, and outcomes. Enables continuous improvement.
Operational data emitted by agents for monitoring and analysis. Feeds evaluation systems.
Immutable records of agent actions and decisions. Supports compliance and accountability.
Managing permissions for agent actions and data access. Fine-grained control reduces risk.
Built-in safety procedures ensuring that if an agent malfunctions, it enters a controlled and safe state.
Predefined alternative paths when agents fail or lack confidence. Protects user experience.
Routing issues to humans or specialized agents when needed. Balances automation with trust.
Automatic detection and recovery from errors or failures. Improves resilience.
Systematic errors in AI outputs caused by skewed training data or assumptions. Can lead to unfair or inaccurate results.
When an AI model's accuracy degrades over time because real-world data changes. Requires continuous monitoring and retraining. ---
Measuring agent performance on defined tasks or benchmarks. Assesses accuracy, efficiency, safety, and user satisfaction.
Comparing agent performance against baselines or alternatives. Informs investment and design decisions.
Testing multiple agent variants to identify optimal performance. Reduces deployment risk.
Incorporating user or system feedback into agent improvement cycles.
Ongoing refinement of agent behavior over time. Supports long-term value.
Resolving customer issues in a single interaction. Key CX metric for agentic systems.
Metric measuring perceived experience quality.
Time it takes for an agent to respond or act. Critical for customer-facing use cases.
Number of tasks an agent system can handle over time. Reflects scalability.
How quickly an organization realizes benefits from deployment.
A structured collection of test cases, rubrics, and scoring criteria used to systematically measure agent performance. Modern evals include task completion, safety, latency, and cost metrics. Examples: MMLU, HumanEval, SWE-bench, GAIA.
Using a language model to evaluate another model's outputs. Scales subjective quality assessment (tone, accuracy, helpfulness) without manual human review. Requires careful calibration to avoid bias.
Assessing the entire sequence of agent actions (not just the final answer) to ensure the agent took reasonable steps. Critical for agentic systems where process matters as much as outcome.
Testing agents for harmful outputs, jailbreak resistance, data leakage, and unauthorized actions before deployment. Includes red teaming, adversarial prompts, and boundary testing.
Re-running agent evaluations after changes to ensure new versions don't degrade performance on previously passing tasks. ---
Software frameworks used to build and manage agents. Choice impacts flexibility, scalability, and vendor lock-in.
Standardizing on a single system to deploy, monitor, and govern agents instead of stitching together point tools.
A limited deployment used to validate value before scaling.
Stability, security, and scalability required for live agent environments.
Meeting requirements for reliability, governance, security, and scale in production.
Dependency on a single technology provider. Buyers should evaluate exit strategies.
The ability for agents to work across systems, frameworks, and vendors. ---
Agents initiating actions without explicit prompts. Unlocks new efficiencies.
Assisting humans with insights while leaving final decisions to people.
Artificially generated data mimicking real-world information. Used for testing and training without exposing sensitive data.
Testing agent behavior in controlled environments before production deployment.
Decentralized training method allowing multiple agents to contribute to a shared model without exchanging raw data.
Seamless hand-off between human operators and agents (e.g., pilot overriding autopilot).
Synthetic media generated by AI to impersonate individuals. Used in misinformation and fraud.
Creating convincing audio that mimics someone's voice. Used in vishing scams and authentication bypass.
Inputs specifically crafted to confuse AI systems. Slight alterations cause misclassification.
Reconstructing sensitive training data by probing a model outputs.
A software development approach where developers describe desired functionality in natural language and let AI agents generate, test, and debug the code. The developer reviews and steers but writes minimal code directly. Popularized by Andrej Karpathy in 2025.
Integrated development environments with deeply integrated AI agents that can read/write code, run terminals, search codebases, and make autonomous edits. Examples: Cursor, Windsurf, Zed. Distinct from simple autocomplete plugins.
An AI agent running on cloud infrastructure rather than locally, enabling persistent sessions, longer execution times, and access to cloud services. Examples: GitHub Copilot Workspace, Devin, Cursor Cloud Agents.
An AI agent running on a user's local machine, with access to local files, tools, and environment. Offers privacy and low latency but limited compute.
Software designed from the ground up with AI as a core capability rather than bolting AI onto existing features. Characterized by conversational interfaces, adaptive behavior, and continuous learning.
A lightweight classifier that routes queries to the most appropriate model (small vs large, general vs specialized) based on query complexity. Optimizes cost and latency.
An architecture where multiple LLMs each process a query, and an aggregator model synthesizes the best elements from each response. Combines diverse model strengths.
The cost and resource management of LLM API calls, including token pricing, caching strategies, and budget allocation across agent workflows.
Storing pre-computed KV caches for repeated prompt prefixes to reduce latency and cost on subsequent requests. Supported by OpenAI, Anthropic, and Google.
A model architecture where different expert sub-networks specialize in different types of inputs, with a gating network routing each token to the most relevant experts. Enables large model capacity with lower per-query compute.
A model that processes both images and text, enabling multimodal reasoning. Examples: GPT-4o, Claude 3.5 Sonnet, Gemini 2.5 Pro. Critical for computer use agents.
The ability to process and reason over very long inputs (100K-1M+ tokens). Enables analyzing entire codebases, books, or conversation histories in a single pass.
The field of research focused on ensuring AI systems behave as intended, avoid harm, and remain under human control. Covers alignment, robustness, interpretability, and governance.
The performance cost of making AI systems safe and aligned with human values. Balancing capability with safety is a core trade-off in AI development.
The gap between what an AI system can already do and what it is currently being used for. As tooling improves, previously latent capabilities become accessible.
The theoretical risk of training future AI models on AI-generated data, leading to progressive degradation of quality and diversity. Requires careful data curation.
A positive feedback loop where AI generates synthetic data used to train better AI, which generates better synthetic data. Enables scaling without proportional increases in human-labeled data.
Methods for aligning pre-trained models to specific behaviors. RLHF uses reward models; DPO skips the reward model; GRPO compares groups of outputs directly. Enables customization for domain-specific tasks. ---