What Is an AI Agent?

From answering to acting

A plain language model is a function: text goes in, text comes out. Ask it to book a flight and the best it can do is describe how you might book one. Nothing changes in the world, and the model never learns whether its advice worked.

An AI agent wraps that same model in a system that lets it act. Anthropic's widely cited definition, from its December 2024 guide "Building effective agents", is a good starting point: agents are systems where the model dynamically directs its own process and tool usage, keeping control over how it accomplishes the task. The important word is directs. The model is not just filling in one step of a pipeline someone else designed; it is choosing the steps.

That one shift, from producing an answer to choosing actions, changes almost everything about how you build, evaluate and secure the system. It is why the rest of this course exists.

The four ingredients: model, tools, loop, environment

Strip away the branding and almost every agent in 2026 is built from the same four parts. Learning to see them separately makes any new agent product much easier to understand.

The environment is the ingredient people most often forget, and it matters a great deal. A coding agent is powerful largely because its environment gives clear, honest feedback: the tests pass or they don't, the compiler accepts the code or it doesn't. Agents struggle most in environments where feedback is slow, ambiguous or missing.

Around these four parts sits a fifth thing that ties them together: the context, the text the model actually sees on each turn. Instructions, tool definitions, conversation history and tool results all compete for space in it. Managing that space well is the subject of Module 3.

goal ──► [ model ] ──decides──► tool call ──► [ environment ]
              ▲                                      │
              └────────── observation ◄──────────────┘
                     (repeat until done)

Workflows vs agents

Anthropic's "Building effective agents" draws a distinction that has become standard vocabulary. Both kinds of system use models and tools; the difference is who decides the control flow.

• Workflows orchestrate models and tools through predefined code paths. A developer decides the steps: first classify the ticket, then draft a reply, then check it against policy. The model fills in each step but does not choose the route.
• Agents let the model decide the path itself, step by step, based on what it observes. The developer supplies tools and a goal; the model figures out the sequence.

The guide's central advice still holds in 2026: find the simplest solution that works, and only add autonomy when it clearly pays off. Many production systems that are marketed as agents are really well-designed workflows, and that is often the right call. Agents trade predictability and cost for flexibility. That trade is worth making when the task genuinely needs it, such as debugging, research, or multi-step operations in messy environments.

How we got here: 2023 to 2026

Agents did not arrive all at once. Each stage below added one capability that the next stage depended on.

1. 1

   Chat assistants (2022–2023)

   General-purpose chat models showed that language models could follow instructions across almost any domain. Early experiments wired them into loops, but the models were too unreliable at multi-step work for those loops to get far.

2. 2

   Tool use becomes a first-class API (2023–2024)

   Model providers added structured function calling: the model returns a JSON request for a named tool instead of prose. In late 2024 Anthropic released computer use (October) and the Model Context Protocol (November), and published "Building effective agents" (December).

3. 3

   Coding agents go mainstream (2025)

   Coding was the first domain where agents became everyday tools, because the environment gives fast, honest feedback. Claude Code previewed in February 2025 and reached general availability in May; OpenAI's Operator (January) was folded into ChatGPT agent in July. Google announced the A2A agent-to-agent protocol in April.

4. 4

   Long-horizon agents (2026)

   Agents began working for hours at a stretch. Hosted agent platforms such as Claude Managed Agents (April 2026) and always-on agents with their own cloud computers (announced at OpenAI DevDay in September 2026) turned agents from a feature into infrastructure.

The research organization METR gives the clearest single measure of this progress. Its time horizon metric asks: how long a software task, measured in skilled-human time, can an agent complete with 50% reliability? The original March 2025 paper found the horizon had been doubling about every seven months since 2019. More recent measurements suggest a doubling closer to every 3.5 to 4 months, roughly 10x per year.

The 2026 landscape

As of October 2026, agents are built on a handful of frontier model families plus a fast-moving open-weight tier. Rankings shift with every release and depend heavily on which benchmark and harness you look at, so this course does not name a single "best" model. Module 10 explains how to read agent benchmarks.

The bigger story of the past year is not models but protocols, and where they live. In December 2025 the Linux Foundation formed the Agentic AI Foundation (AAIF). Its founding projects were Anthropic's Model Context Protocol (MCP), Block's open-source goose agent, and OpenAI's AGENTS.md instruction-file convention. Google's A2A protocol, which had already moved to the Linux Foundation in mid-2025, joined AAIF as a Growth Stage project on August 27, 2026, keeping its own technical steering committee.

Having competitors' protocols under one neutral foundation matters for builders. It means a tool server you write once can be used by agents from different vendors, and an agent you build can hand work to agents built on entirely different stacks.

Why agent-to-agent matters

The theme running through this course, and through aikind.com, is that the next phase of agentic AI is about agents working with other agents. A useful shorthand that took hold in 2026 is:

• MCP is vertical. It connects one agent down to the tools and data it controls. A tool is a deterministic function: you call it, it returns a result.
• A2A is horizontal. It connects an agent across to other agents it does not control. The other agent is opaque: it has its own model, memory, tools and judgment, and it may take hours to finish.

Why would you want agents to talk to agents at all, instead of giving one agent every tool? Three reasons come up again and again.

1. 1

   Specialization

   An agent tuned for tax law, or for your company's codebase, will often outperform a generalist that has the same tools. Delegating to specialists is how human organizations work too.

2. 2

   Organizational boundaries

   Your travel agent can't be given your airline's internal database credentials, but it can ask the airline's agent to rebook a flight. A2A lets agents cooperate without exposing their internals to each other.

3. 3

   Context limits

   Every agent has a finite context window. Splitting a large job across agents, each with its own clean context, is one of the most effective ways to scale. Module 6 covers when this helps and when it backfires.

Agent-to-agent communication also raises new questions: how does one agent verify another's identity, how do agents pay each other, and who is accountable when a chain of agents does something wrong? Track 2 of this course (Modules 5 to 8) takes those questions one at a time.