New AI terminology piles up faster than most people can absorb it. Multimodal, agent, tool calling, MCP — these words show up in every product pitch, yet the definitions rarely stick, because they aren't parallel trivia to memorize one by one. They describe three layers of the same machine: what it can take in, whether it decides things on its own, and how it gets things done. Once those layers click, the next new term has an obvious home.
Four foundational terms come first. Artificial intelligence (AI) is the outermost circle: any effort to get a machine to perform tasks that once required human intelligence. Machine learning is one path inside it — instead of hard-coding every rule, the system learns patterns from data. Deep learning is a technique within machine learning: finding those patterns with multi-layer neural networks. Nearly every large model today runs on deep learning.
AI, machine learning, and deep learning form a clean, uncontested containment relationship. "Large language model (LLM)" is different: most LLMs use deep learning, but "LLM" isn't a precisely bounded technical category, so that layer only belongs on the diagram as a dotted line, not a solid one.
One myth worth clearing up: "neural network" borrows its name from biology, but it's really a stack of computational layers, and its "parameters" aren't neurons — just numeric values tuned during training. More of them doesn't mean a smarter model.
Multimodal, agent, tool calling, and MCP don't slot into these circles at all — they're three separate things commonly paired with large models, each answering a distinct question.
Send an AI a photo and ask what it says, and it answers in seconds — looking like it "saw" the whole image. What's actually happening is tiling: the model receives chopped-up patches, each measured like text is — in tokens, the base unit models use to count content.
Multimodal describes a model that can take in or produce more than text — images and audio too. Accepting and producing are separate questions, and an uploaded image gets sliced into patches that each cost tokens — major vendors' docs describe broadly the same process, with only patch size differing — so a larger, higher-resolution image burns more.
Being able to read an image doesn't mean a model can draw one. As of August 2026, Claude's FAQ page says it plainly: it understands and analyzes images but cannot generate, edit, or create them — a recognized multimodal model that still can't produce a picture. The reverse doesn't hold either: some models can generate images or speech, so input and output must be checked separately.
Multimodal also doesn't guarantee accurate reading. A model estimates object counts only roughly, less reliably for small or numerous items — a limitation Anthropic's documentation admits, alongside its inability to judge whether an image is AI-generated. Any conclusion an AI draws from a picture still needs a human check. Multimodal isn't a model category; it's a property describing what a model takes in and out.
Imagine three AI-powered products: one follows a sequence a human wrote out in advance; one chats and advises but never acts, leaving the task to a human; one decides its own order with no one telling it what's next. Only the third genuinely qualifies as an agent.
Definitions of "agent" look scattered until one distinction separates them: who decides each step. In a December 2024 engineering post, Anthropic drew this line: when an LLM and tools are orchestrated through predefined code paths, that's a workflow; when the LLM dynamically directs its own process and tool usage, that's an agent. Alibaba Cloud's Bailian documentation states the identical criterion: a workflow is precisely controlled by a predefined process, while an agent has the model autonomously plan its own steps.
Three points hold across most definitions: an agent must call tools (files, commands, a database, an API — talking alone doesn't count); it typically handles multi-step tasks; and a chatbot alone doesn't qualify.
"Agent" also carries an older meaning: in control theory and classical AI, it's anything that perceives its environment and acts on it — a thermostat qualifies. Today's coverage means the newer sense: LLM-driven, deciding its own next step. Tellingly, Anthropic's own glossary doesn't even list "agent" as a term.
Agent products advertise "checks the weather," "sends the email," "tracks your package," as if the model does these things itself. But a model is fundamentally a program that processes content — it reads whatever is handed to it, nothing more. It isn't connected to the internet and can't reach your inbox.
The real mechanism is tool calling (also called function calling), and the model never executes anything. It outputs a message specifying which tool to use and what to fill in; code outside the model runs it and feeds the result back. A full round trip has four steps — model requests, external program executes, result returns, model responds — and only the first and last are the model's own work.
Vendors describe this identically: Alibaba Cloud's Bailian docs say running the tool function happens in the user's own computing environment, not inside the LLM; Anthropic's docs say the model only issues a structured request that the user's code (or Anthropic's servers) carries out, with the result returned to the conversation; Google and OpenAI describe the same flow with a different step count. So when an AI has "searched the web," an external program — on the user's machine or the vendor's cloud — ran the search and fed results back; sending an email depends on whether that layer has account access wired up.
One confusing detail: the information filled in is also called a parameter — not a "model parameter," fixed during training. Even the terminology hasn't settled: "tool calling" and "function calling" mean the same thing.
The model side is sorted, but the tool side has a problem: every new tool traditionally needed its own custom integration, and switching AI applications meant starting over.
The standard built to fix this is the Model Context Protocol (MCP): shared rules for naming and passing information so independent implementations work together. A tool describes itself once, and any MCP-supporting application can discover and connect to it. MCP's own site compares it to a USB-C port: instead of every device needing its own cable, one shared port works everywhere.
The most important thing about MCP is who governs it. As of August 2026, no single company maintains it — in December 2025, stewardship moved to a new foundation under the Linux Foundation, jointly overseen by several competing companies, which says more about its industry standing than any adopter list would.
One point easy to get backwards: MCP does not replace tool calling. A tool is first described, discovered, and connected under MCP's rules; the call itself then works exactly like the tool calling above, and whether to call it remains entirely the model's decision. MCP's own site draws this boundary explicitly: it governs context exchange, not how an application uses the model, so "MCP replaces tool calling" contradicts its own positioning. It fills the gap around discovery and connection, not the model's decision to use a tool — you can build an agent without MCP, just with each integration done separately.
Does sending an image to AI use up tokens? Yes. It's sliced into patches that each count as tokens like text does — larger, higher-resolution images produce more patches and more tokens.
If a model can understand images, can it also generate them? Is it reliable at counting objects or spotting fakes? Not necessarily — accepting and producing images are independent capabilities. Claude's FAQ page says it only understands images and cannot generate, edit, or create them. Treat its visual conclusions as reference only: object counts are rough estimates, and it cannot reliably tell whether an image is AI-generated.
Can you build an agent without MCP? Yes. The core mechanism — model requests, external program executes — predates MCP and worked fine with each vendor's own interface, just without standardization.
How do you tell whether a self-described "agent" actually is one? Check who decides the next step. A fixed path needing a code change for every new scenario is a workflow; only autonomous decision-making plus tool calling meets the bar, regardless of tool count or which LLM powers it.
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 9 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 9, free | Last updated: 2026-08-28
