AI Agent Glossary: Every Term You Need to Know

The core mechanism of an AI agent: perceive the world, decide what to do, take action, observe the result, repeat. Each iteration of the loop moves the agent closer to completing its task (or it gets stuck and asks for help). The agentic loop is what distinguishes agents from chatbots — chatbots don't have a loop, they just generate text in response to a single prompt.

An AI system that takes actions on a user's behalf — browsing the web, clicking buttons, sending emails, editing files — without step-by-step human instruction. Agents combine a large language model with a tool-use loop and (usually) some form of memory. By 2026, "agent" specifically means a system that can close the loop on real tasks, not just generate text.

A standardized way for software to communicate with other software. Agents use APIs to interact with external services — sending email via the Gmail API, posting to Slack via the Slack API, etc. The breadth and quality of available APIs is one of the main things that determines what an agent can do.

The degree to which an agent can operate without human intervention. "Fully autonomous" means the agent completes its task with no human input; "human-in-the-loop" means a human approves critical actions. Most production agents in 2026 are partially autonomous — they handle routine steps but ask for confirmation on high-stakes actions like payments or deletions.

An agent that operates by controlling a web browser — clicking, typing, and navigating web pages. Browser agents like OpenAI Operator and Google Mariner are specialized for web tasks and can't control native desktop apps. They're easier to make safe because the browser sandbox limits what they can affect.

An AI system that generates text in response to user prompts. ChatGPT, Claude, and Gemini are chatbots. Unlike agents, chatbots can't take actions in the world — they can only suggest actions for the user to take. See our Agents vs Chatbots guide for a detailed comparison.

A specific category of agent that controls a computer by moving the cursor, clicking, and typing — essentially simulating human interaction with a desktop. Anthropic's Claude Computer Use is the canonical example. Computer use agents can control any application a human could, but pose greater safety risks than browser-only agents.

The amount of text an LLM can hold in working memory at one time, measured in tokens. Modern LLMs have context windows of 100,000 to 2,000,000 tokens. Larger context windows let agents handle longer tasks but increase cost and latency. When an agent runs out of context, it has to summarize or forget earlier steps.

An agent that controls native desktop applications rather than (or in addition to) web browsers. Claude Computer Use is the leading example. Desktop agents are more powerful than browser-only agents but require more careful safety configuration.

A numerical representation of text that captures its semantic meaning. Embeddings let agents compare pieces of text for similarity, which is the basis for semantic search and memory retrieval. When an agent "remembers" a past conversation, it's typically finding similar embeddings in a vector database.

Providing an LLM with a few examples of the task you want it to perform, as part of the prompt. Few-shot prompting dramatically improves agent accuracy on specific tasks. Most production agent deployments use few-shot examples to teach the agent the desired behavior.

A capability of modern LLMs to output structured instructions for calling external functions (tools) rather than just text. Function calling is the technical mechanism that lets agents take actions — the LLM emits a function call, the surrounding software executes it, and the result is fed back into the LLM.

Safety mechanisms built into or around an agent to prevent it from taking harmful actions. Guardrails include confirmation prompts for sensitive actions, deny-lists of apps or websites the agent can't touch, step limits to prevent runaway loops, and content filters to prevent harmful outputs.

When an LLM confidently asserts something that isn't true. In an agent context, hallucination can mean clicking a button that doesn't exist, calling an API with invalid arguments, or reporting a "success" when an action actually failed. Modern agents mitigate hallucination with verification steps, but it can't be eliminated entirely.

A design pattern where a human approves critical actions before the agent takes them. HITL is the standard safety mechanism for high-stakes agent workflows — payments, deletions, external communications. The trade-off is speed: every approval step adds latency. Most production agents use HITL selectively rather than for every action.

A training process that teaches an LLM to follow instructions and break down tasks into steps. Instruction-tuned models are essential for agents — without this training, the model would generate reasonable-sounding text but not know how to use tools or structure actions.

The underlying AI model that powers modern agents. LLMs are neural networks trained on billions of pages of text and code, giving them a working knowledge of language, reasoning, and the world. GPT, Claude, and Gemini are all LLMs. Every modern agent uses an LLM as its "brain."

An agent's ability to retain information across steps or sessions. Short-term memory is the LLM's context window — what it can hold in working memory during a single task. Long-term memory is typically a vector database that stores past interactions and lets the agent retrieve relevant ones. Memory is one of the hardest problems in agent design.

A system where multiple specialized agents collaborate on a task, each handling a different aspect. Relevance AI and Lindy.ai both support multi-agent workflows. The advantage is specialization — each agent can be optimized for its specific role. The disadvantage is complexity in coordination.

A platform design that lets non-technical users build agents through a visual interface, without writing code. Lindy.ai is the leading no-code agent platform. The trade-off is usually flexibility — no-code platforms are easier to use but less customizable than low-code or code-first alternatives.

The coordination of multiple tools, agents, or steps to accomplish a complex task. An orchestration layer decides which agent or tool to invoke when, handles handoffs between them, and manages the overall workflow. Modern agent platforms (Lindy, Relevance, Copilot Studio) are essentially orchestration layers.

How an agent receives information about the world. Perception can be through screenshots (for computer-use agents), API responses (for tool-using agents), or structured data (for database-querying agents). The perception method shapes what an agent can do — screenshot-based agents are more flexible but slower and more error-prone.

The instruction given to an LLM to elicit a response. In an agent context, prompts are often structured to include the current state, available tools, and the task to accomplish. Prompt engineering — crafting effective prompts — is a key skill for getting good results from agents.

A technique where an agent retrieves relevant information from an external knowledge base before generating a response. RAG is what lets agents answer questions about your specific data — your documents, your CRM, your codebase — rather than just their training data. Most production agents use some form of RAG.

An LLM's ability to break down a problem into steps and work through them logically. Modern "reasoning models" (like OpenAI's o-series and Anthropic's Claude with extended thinking) are explicitly trained to reason before responding, which improves agent performance on complex tasks.

The combination of technical mechanisms and operational practices that prevent agents from causing harm. Safety includes guardrails (technical), permission models (configuration), audit logging (monitoring), and human-in-the-loop oversight (operational). See our AI Agent Safety Guide for a detailed treatment.

An agent perception method where the agent receives screenshots of the screen (or browser) at each step. This is how computer-use agents like Claude Computer Use work. The advantage is flexibility — any UI a human can use, the agent can use. The disadvantage is latency and error-proneness.

An agent's ability to call external tools — APIs, file operations, browser actions — to take actions in the world. Tool use is what makes an LLM into an agent. Without tools, the LLM can only generate text; with tools, it can do almost anything software can do.

The full sequence of perceptions, decisions, and actions an agent takes to complete a task. The trajectory is the agent's "execution log" and is invaluable for debugging, auditing, and improving agent behavior. Production agents always log their trajectory for later review.

The event that starts an agent workflow. Triggers can be time-based ("every Monday at 9am"), event-based ("when a new email arrives"), or manual ("when I click this button"). Most production agent platforms support all three trigger types.

A specialized database for storing and retrieving embeddings (numerical representations of text). Vector databases are the technical foundation of agent long-term memory and RAG systems. Popular vector databases in 2026 include Pinecone, Weaviate, and pgvector (the Postgres extension).

A defined sequence of steps an agent takes to accomplish a task. Workflows can be simple ("check the weather and tell me if I need an umbrella") or complex ("research 5 competitors, draft a competitive brief, schedule a meeting to review it"). Most agent platforms let you save workflows as templates for reuse.

A marketing term used by platforms like Relevance AI and Lindy.ai to describe a collection of specialized agents that work together like a team of employees. The metaphor is useful but imperfect — agents don't have the judgment, relationships, or adaptability of human employees. They're best understood as a new category of automation tool, not as actual employees.