AI agent glossary: plain-English terms for founders
Short, plain definitions of the AI agent terms founders and small teams run into, with the differences that actually matter.
An AI agent is software that uses an AI model to plan and take actions toward a goal, choosing its own next steps and tools; a workflow follows steps someone defined in advance.
What is an AI agent?
An AI agent is software that uses an AI model to work toward a goal by deciding its own next steps, using tools such as search, files or other apps, and checking the results as it goes. You describe the outcome; the agent works out the route, within the limits you set.
Anthropic draws the line this way: workflows are "orchestrated through predefined code paths," while agents "dynamically direct their own processes and tool usage." Many useful setups mix both.
How is an AI agent different from a workflow, a chatbot and an assistant?
The difference is who chooses the steps and whether the software takes actions. The terms are used loosely, so read the table as common meaning, not a standard.
| Term | Who chooses the steps | Takes actions? | Typical use |
|---|---|---|---|
| Workflow | A person, in advance | Yes, the ones it was built to take | Repeating the same process reliably |
| Chatbot | A person, one message at a time | Usually not; it answers | Questions and conversation |
| Assistant | Mostly the person, with suggestions | Sometimes, when asked | Helping one person with tasks |
| AI agent | The agent, toward a goal you set | Yes, with tools | Multi-step work with some independence |
What are the core terms?
Model (LLM)
The AI system that reads text and produces text, such as a large language model (LLM). It's the reasoning engine inside an agent. On its own, a model only answers; the agent around it gives it tools, memory and a loop so it can act.
Prompt and system prompt
A prompt is any instruction you give a model. A system prompt is a standing set of instructions that applies to every exchange, such as the agent's role, rules and tone. Task-specific instructions usually go in a brief.
Tool and tool call
A tool is a capability an agent can use, such as web search, reading a file or sending a message. A tool call is the moment the agent uses one. Tools are what turn a model that writes into an agent that does.
Agent loop
The cycle an agent repeats until the task is done: look at the situation, decide what to do, act, then check the result. Google's glossary describes the stages as observe, reason, act and feedback, repeated until a stopping condition is met.
Context window
The amount of text a model can consider at once, including instructions, files and the conversation so far. Anything outside it is invisible to the model in that moment, which is why good briefs give the right context rather than all of it.
Memory
How an agent keeps information beyond a single exchange, for example notes saved to files or a database it can search later. Memory is commonly used to mean anything that persists between runs, so ask what a tool actually stores and where.
MCP (Model Context Protocol)
An open standard for connecting AI applications to tools and data. Its documentation calls it "an open-source standard for connecting AI applications to external systems" and compares it to "a USB-C port for AI applications": one plug shape instead of a custom cable for each tool.
Skill
Commonly used to mean a packaged set of instructions, and sometimes files or scripts, that an agent can load for a particular kind of task, such as writing a report in your format. A skill tells the agent how to do something; a tool lets it act.
What terms describe several agents working together?
Multi-agent system
A setup where several agents work on related tasks, often with different roles. It can cover more ground in parallel, and it needs clear boundaries so agents don't duplicate work or leave gaps between them.
Orchestrator (manager agent)
An agent that splits a goal into tasks and assigns them to other agents. Google's glossary defines a manager agent as "an agent that controls one or more sub-agents." In a human-led company, the person still sets the goal the orchestrator works toward.
Subagent
An agent that another agent hands a specific piece of work to. Subagents usually have a narrower job and fewer tools. In some tools, each subagent runs in its own context window with its own instructions and permissions, then returns a summary.
Handoff
The moment work passes from one agent, or person, to another, and the note that goes with it: what's finished, what's open and what to watch. A good handoff note is short, saved next to the work and readable by whoever picks it up.
Evaluator (reviewer) agent
An agent that checks another agent's work before it's finalized. Google's glossary describes one agent "manufacturing a product and a separate agent… testing that product before it is released." The agent that did the work shouldn't be the one that signs it off.
Orchestration
The coordination of tasks across several agents or model calls: breaking work down, routing it and combining the results. It's commonly used to describe the system as a whole rather than any one agent.
What terms describe control and quality?
Human in the loop
A person approves certain agent actions before they happen. The agent pauses at the steps you've marked as sensitive and waits. The hard part isn't the term; it's choosing which actions. Our checklist on which AI agent actions need human approval covers that.
Approval gate
A specific checkpoint where work stops until a person says yes, such as before sending an email, spending money or deleting a file. Gates keep high-stakes actions with a person and let everything else keep moving.
Guardrail
An automatic check on what an agent receives, produces or does, such as blocking a disallowed request or validating output before it leaves the system. Guardrails run without a person; approval gates wait for one.
Permissions
What an agent is allowed to access or do: which files it can read, which tools it can use, which accounts it can touch. A good default is the smallest set that lets the agent finish its task.
Definition of done
A short list of checks that tells anyone, agent or person, when a task is finished. "Done" should be something you can open and verify, such as a saved file or a working change, not a message that says "finished."
Trace (log)
A record of what an agent did during a run: the steps it took, the tools it called and what came back. Traces help you understand a result after the fact and find where something went wrong.
Eval
A repeatable test of how well an agent or model performs on a set of tasks. Teams use evals to compare setups and to check that a change didn't make results worse.
What terms describe running the work?
Agent harness
The code around a model that runs the agent loop: it passes in instructions, calls tools, keeps track of context and decides when to stop. The term is used loosely; it's commonly used to mean everything that turns a model into a working agent.
Brief
A short written handoff for one task: the goal, why it matters, context, deliverable, definition of done, limits, and when to stop and ask. See our AI agent brief template.
Run
One complete attempt by an agent at a task, from the first instruction to the final result or stop. Asking "what happened in that run?" usually means reading its trace.
AI agent dashboard
A view that shows the work your agents are doing: who's on what, what's done, what's stuck and what needs your decision. See what an AI agent dashboard should show.
Virtual office for AI agents
A shared digital space where AI agents and the people who direct them work in one place. The phrase covers two ideas: an office for a human team that agents join, and an office built around agents. See our explainer on virtual offices for AI agents.
AI-run company
A business where AI agents carry out much of the day-to-day work while a person sets the direction, makes the consequential calls and stays accountable. See what is an AI-run company? for the human-led definition.
Operating system for AI-run companies
Plenarly's term for one place where a founder connects AI agents from any lab, organizes the company's goals, projects and the work those agents carry, approves their work and makes the calls. It describes what Plenarly is being built to be, not a released product.
Where does Plenarly fit?
Plenarly is being built as the operating system for AI-run companies: one place where founders and small teams connect their AI agents from any lab, approve their work, and manage the apps those agents build. Early-access sign-ups are open.
Join early access if you already run AI agents on real work. For plain facts, see What Plenarly is and isn't.
Frequently asked questions
Is "agentic AI" the same as an AI agent?
Not exactly. Google's glossary calls "agentic" "the adjective form of agent," referring to qualities such as autonomy. "Agentic AI" usually describes systems that plan and act toward goals, so the terms overlap heavily. An AI agent is the thing; agentic describes how it behaves.
Do I need to know these terms to use AI agents?
No. You can lead AI agents well with plain language: a clear goal, a definition of done and limits on what needs your approval. The terms help when you compare tools or read documentation, and a handful, such as brief, handoff and human in the loop, go a long way.
What's the difference between an AI agent and a harness?
The agent is the whole working thing that pursues a goal. The harness is the code around the model that makes that possible: it runs the loop, calls tools, manages context and decides when to stop. Model, harness, instructions and tools together add up to an agent.
What's the difference between a skill and a tool?
A tool lets an agent do something, such as search the web or read a file. A skill, as the term is commonly used, teaches the agent how to do a kind of task well, usually as packaged instructions. A reporting skill might call several tools.
Sources
- Anthropic, "Building effective agents," December 19, 2024. Quotes: workflows are "orchestrated through predefined code paths"; agents "dynamically direct their own processes and tool usage." Checked on the live page Oct 24, 2026, ~8:35 PM ET. https://www.anthropic.com/research/building-effective-agents
- Google for Developers, "Machine Learning Glossary: Agentic." Definitions of agentic loop (observe, reason, act, feedback), manager agent ("an agent that controls one or more sub-agents"), evaluator agent, agent orchestration and "agentic" ("the adjective form of agent"). Checked Oct 24, 2026, ~8:35 PM ET. https://developers.google.com/machine-learning/glossary/agent
- Model Context Protocol, "What is the Model Context Protocol (MCP)?" Quotes: "an open-source standard for connecting AI applications to external systems"; "a USB-C port for AI applications." Checked Oct 24, 2026, ~8:35 PM ET (the URL redirects to a dated docs path). https://modelcontextprotocol.io/docs/getting-started/intro
- Anthropic, "Create custom subagents," Claude Code Docs. Background for the subagent entry: each subagent "runs in its own context window with a custom system prompt, specific tool access, and independent permissions." Checked Oct 24, 2026, ~8:35 PM ET. https://code.claude.com/docs/en/sub-agents