The AI Agent Glossary: 10 Terms Explained in Plain English

July 23, 2026 · 8 min read · By bots.team

"Agent," "assistant," "skill," "MCP," "context window" — the AI-tools world throws around a lot of words that sound similar but mean different things. Here's what each one actually means, with a real example for every term.

An AI agent plans and acts toward a goal across multiple steps without you prompting each one; an AI assistant replies only when you ask it something; automation follows fixed steps you defined in advance rather than deciding anything itself. A skill or plugin is one specific capability an agent can use (like "send an email"); MCP is the shared standard that lets different companies' agents and tools talk to each other. A trigger is what starts a bot's work (a schedule or an event); human-in-the-loop means it pauses for your approval before doing something risky; a context window is how much text the underlying model can "see" at once; and tokens or credits are the metered units some tools bill you by, instead of a flat monthly price.

Why this list exists

Nobody who's building a business, running a creator account, or managing social media for a client signed up to also become an AI terminology expert. But the terms matter in a very practical way — they show up on pricing pages, security pages, and comparison charts, and knowing what they mean is often the fastest way to tell whether a tool actually fits what you need, or whether "agent" is being used loosely to describe something closer to a chatbot with extra branding.

These ten terms come up across nearly every AI agent, assistant, or automation tool on the market in 2026 — not just bots.team's. Where a term has a nuance worth knowing before you buy, it's called out below.

The 10 terms

1. AI agent

Software that takes a goal, breaks it into steps, uses tools to act, checks the result, and keeps going — without you approving every individual step. The defining trait is the loop: observe, plan, act, evaluate, repeated until the goal is done or it needs your input.

Example: "Check my competitor's pricing page every morning and tell me if it changes" is an agent job — it runs unattended and only reports back when there's something to say.

2. AI assistant (or chatbot)

A tool that responds to what you type, one exchange at a time, and then waits. It doesn't keep working after you stop typing, and it doesn't act on a schedule without you present.

Example: Asking ChatGPT to draft a caption is an assistant interaction — useful, but it stops the moment you close the tab. The line between "assistant" and "agent" is genuinely blurry in 2026 marketing copy; the test that actually holds up is whether the tool keeps working when you're not there.

3. Automation

Automation follows a fixed sequence you defined step by step in advance — "when X happens, do Y." It doesn't reason about the goal or adapt if something unexpected happens; it just runs the recipe. Zapier-style tools and traditional RPA (robotic process automation) are automation, not agents, even when they're described as "AI-powered."

Example: "When a new row appears in this spreadsheet, send this exact email" is automation. "Watch this spreadsheet and tell me if anything looks off" is agentic — it requires judgment about what "off" means.

4. Skill / plugin / tool

A single specific capability an agent can call on — reading a file, sending an email, browsing a webpage, running a calculation. Agents are typically built from a small set of general reasoning ability plus a larger set of narrow skills they can invoke as needed.

Worth checking: some tools let anyone publish community skills to a shared marketplace. That can be genuinely useful, but it's also a real risk surface if it's not reviewed — OpenClaw's skill marketplace was found to have roughly 12% malicious skills in 2026, including its single most-downloaded one.

5. MCP (Model Context Protocol)

A shared technical standard, originally released by Anthropic, that lets an AI agent connect to outside tools and data sources in a consistent way — instead of every company building its own one-off integration format. MCP was donated to a Linux Foundation-hosted, cross-vendor foundation in 2026 and now has support from Anthropic, OpenAI, Google, and Microsoft's own agent runtime.

Why it matters to you: it's a decent proxy for how "open" a tool's ecosystem is. A tool built on MCP-shaped connections is generally easier to extend or swap components in later than one with entirely proprietary plumbing.

6. Trigger / schedule

What actually starts an agent's work. A schedule runs on a fixed cadence ("every morning at 6am"); a trigger runs in response to an event ("when this file changes," "when this email arrives"). This is the mechanism that lets an agent work without you present — the thing a plain chatbot fundamentally can't do on its own.

7. Human-in-the-loop / approval mode

A safeguard where the agent pauses and asks before doing something hard to undo — deleting a file, sending a message, spending money — instead of just doing it and telling you afterward. Some tools offer several distinct levels: read-only (can look but not touch), approval-required (asks first), and fully autonomous (acts within limits you set once).

Being direct about where bots.team stands: Town, a newer AI-assistant competitor, documents three explicit modes like this. bots.team's approach is a deliberate trade-off, not an oversight — you approve the plan once, then the bot runs unattended on schedule, since a bot that stops to ask mid-run isn't really unattended anymore. Worth knowing if per-action approval matters more to you than unattended scheduling.

8. Context window

The amount of text (measured in tokens — see below) a model can "see" and reason over at one time, including your instructions, any files it's reading, and its own prior steps in a task. A bigger context window means an agent can work through a longer document or a longer chain of steps without losing track of earlier details.

9. Token / credit

A token is a small chunk of text (roughly ¾ of a word) that AI models process and get billed by internally. A credit is a company's own metered billing unit, often built on top of token usage, used to charge per-task instead of a flat monthly fee. These are the units behind "usage-based pricing" — Microsoft's Copilot Cowork bills $0.01 per credit (a heavy task can run $7+), and Town bills $0.03–0.044 per credit past its plan allowance.

Why it's worth knowing: "credits" sounds friendlier than "metered billing," but it's the same idea — your bill scales with how much you actually use the tool, which can be a feature (you only pay for what you use) or a surprise (a busy month costs more), depending on how predictable you need your budget to be.

10. Headless

Running a tool without its normal visual interface — no chat window, no browser tab open, just the underlying engine executing a task in the background and reporting a result. This is the technical shape behind most scheduled AI agent products, including bots.team: each bot is a headless session of Anthropic's Claude Code, running on a timer or a trigger rather than a screen someone is watching.

Quick reference: which term applies to which tool?

One honest note on "flat pricing": bots.team's $0/$5/$19 tiers cover the app itself — scheduling, standing plans, and reporting. Running a bot also requires your own Claude subscription (Claude Pro starts at $20/month), so the realistic all-in cost is roughly $20–$39/month, not $0–$19 alone. Both fees are flat and capped, which is still more predictable than a per-task credit meter — but it's two numbers, not one.

FAQ

Is "AI agent" just a rebrand of automation? No — the practical difference is real. Automation follows a rule you wrote ("if this, then that"). An agent is given a goal and figures out the steps itself, which means it can handle situations you didn't explicitly plan for. That flexibility is also why agents need more thought put into approval and safety controls than a simple automation does.

Does every "AI agent" tool actually behave like an agent? Not always. Some products use "agent" as a marketing label for what's really a well-designed chatbot with a few extra buttons. The test that holds up: does it keep working on a schedule or trigger without you present, or does it only respond when you type something? If it's the second one, it's an assistant.

Why does it matter whether a tool uses MCP? It doesn't change what the tool does for you today, but it's a signal about flexibility later — MCP-based tools are generally easier to connect to new services as they emerge, since they're not locked into one company's proprietary format.

Are tokens and credits the same thing? Related but not identical. A token is the underlying unit AI models process text in. A credit is a company's own billing wrapper, often priced per task rather than per token directly, which is why a "medium" task and a "heavy" task can cost different numbers of credits even though both are ultimately built from tokens under the hood.

Once the terms stop being a barrier, the actual decision is simpler than it sounds: do you want something that answers when asked, or something that keeps working on a schedule and only speaks up when there's something to report? bots.team is built specifically for the second kind of job — describe it once in plain English, and it runs standing, scheduled, flat-priced, with no inbound network door left open for anyone to find.

Sources: Anthropic's Model Context Protocol documentation and Linux Foundation donation announcement, Microsoft's 365 Blog Copilot Cowork pricing, Town AI independent reviews, OpenClaw's ClawHub marketplace disclosures, and bots.team's own product and pricing pages. Figures current as of July 2026.