AI Agents Overview
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
An AI agent is an assistant that has been given a goal, tools, and permission to keep working until the goal is met. In 2026 that is no longer an exotic setup. The same Claude, ChatGPT, Gemini, or Copilot subscription you already use has an agent mode built in, and the coding tools most developers use every day are agents by default.
This track teaches you to delegate to agents well: what to hand over, how to brief them, how to keep them inside safe boundaries, and how to tell whether they actually did the job. It is the flagship of the Build section, and it contains two sub-tracks, MCP and Loops, for readers who want to go deeper.
What an agent is in 2026
The shift from assistant to agent is a shift in who does the work. With an assistant, you ask a question and get an answer, then you carry out the steps yourself. With an agent, you describe the outcome, point it at your files and connected apps, approve the risky steps, and check the result.
Instead of asking:
"Write me a Python script to analyze this CSV"
You tell an agent:
"Analyze [sales_data.csv], identify the three biggest trends, create charts for each, and draft a one-page summary for [the leadership meeting on Friday]. Ask me before sending anything."
The agent reads the file, writes and runs the analysis, produces the charts, drafts the summary, and stops at the approval point you set. Every agent you will meet in this track is some version of that pattern: a model, a set of tools, a loop that keeps going until a stop condition is met, and a human who decides what needs a yes.
What changed in 2026 is where agents live. Claude folded its Cowork agent into the main app. ChatGPT added ChatGPT Work and always-on Dots, and Anthropic answered in October with the Claude Duet preview. Gemini and Microsoft Copilot gained agent modes. Standalone agent products, including the AI browsers launched in 2025, were mostly retired and absorbed into the assistants people already pay for. You do not need to buy an agent; you need to learn to use the one you have.
Who this track is for
Three kinds of readers come here, and the track serves each of them with a different path.
Everyday professionals who want to hand real work to the agent inside their chat app and not get burned. You need pages 2, 3, 4, 13, and 14, and about ninety minutes.
Builders who want to create an agent for a specific job, whether that is a no-code scheduled task or a Python loop against the API. You need pages 5, 6, 8, 9, the MCP sub-track, and page 12.
Practitioners who run coding agents and agent systems for a living and want the current vocabulary: subagents, agent teams, harnesses, hooks, evals, containment. You need page 7, the Loops sub-track, and pages 6, 12, and 13.
If you have never used an AI assistant for anything beyond chat, start with the AI Fluency track first. This track assumes you can already write a clear brief and check an answer.
How the track is organized
The pages below appear in the sidebar in this order. Positions 10 and 11 are sub-tracks with their own pages.
- AI Agents Overview (this page)
- Agents vs Assistants: the four properties that make something an agent, and when a plain assistant is the better choice
- Agents in the Apps You Already Pay For: Claude, ChatGPT Work and Dots, Gemini, Copilot, and Muse, with the three controls that matter: approvals, scoping, and blast radius
- The 2026 Agent Landscape: the quarterly state-of-agents page, with adoption numbers, categories, standards, and what to watch
- Build Your First Agent: one tutorial, two paths: a no-code scheduled agent and a Python agent loop against the API
- Patterns and Multi-Agent Systems: the agent loop, planning and evaluation, and the four multi-agent shapes, mapped to subagents, agent teams, and workflows
- Claude Code: Your AI Coding Agent: surfaces, Plan Mode and auto mode, CLAUDE.md, skills, subagents,
/loopand/goal, Routines, hooks, and Code Review - Skills, AGENTS.md, and CLAUDE.md: the Agent Skills open standard and the two project-instruction files every agent reads
- Customize: Projects, Skills, RAG, Fine-Tuning: a short decision page for when a brief is not enough
- MCP Explained: Connectors for Users, Protocol for Builders: the MCP sub-track, from connecting your first server to building and securing one on the 2026-07-28 spec
- Loops and Harnesses: the Loops sub-track, from the anatomy of an agent loop to the overnight coding recipe and when a workflow beats a loop
- Evals for Agents: tasks, trials, graders, pass@k versus pass^k, and how to start with twenty tasks drawn from real failures
- Agent Safety and Guardrails: the 2026 incident record and the safety posture that follows from it
- 50+ Agent Use Cases: a catalog of recipes across software, marketing, support, finance, HR, and more
Three paths through the track
The delegation path (about ninety minutes). Read Agents vs Assistants, then Agents in the Apps You Already Pay For, then Agent Safety and Guardrails. Finish by picking one recipe from the use-case catalog and running it inside your own app with approvals turned on. You will come away able to hand a multi-step task to an agent, scope what it can touch, and know what to check afterwards.
The builder path (an afternoon). Read Build Your First Agent and complete the no-code path. Then read Skills, AGENTS.md, and CLAUDE.md and turn your brief into a reusable skill. Read the Customize page to decide whether you need retrieval or fine-tuning (you almost certainly do not yet). Connect your first MCP server in the MCP sub-track, and finish with Evals for Agents so you have twenty tasks that tell you whether the agent is getting better.
The practitioner path (a week of evenings). Start with the Claude Code guide for the current primitives. Work through the Loops sub-track to understand stop conditions, budgets, and harness design. Return to Patterns and Multi-Agent Systems for subagents versus agent teams versus workflows, then build a regression eval and a containment plan with the last two pages.
What you will be able to do
By the end of this track you will be able to decide whether a task belongs with an agent, a workflow, or a person; write a brief an agent can act on without you in the room; grant tools and connectors with the minimum reach the job needs; read what an agent did and verify the result; and recognize the three failure classes of 2026 (over-privileged keys, injected instructions, and runaway spend) before they happen to you.
A note on vocabulary
The words you will meet most often are agent (a model that plans, acts, and loops toward a goal), tool or connector (a capability the agent can call, usually exposed through MCP), approval (a step that waits for your yes), routine or scheduled task (an agent that runs on a timer without you), blast radius (everything an agent could change or delete if it went wrong), harness (the scaffolding of prompts, agents, tools, and orchestration around a model), and eval (a repeatable test of whether an agent does a task correctly). Each is defined where it first matters, and the AI Glossary collects them.
Key takeaways
- An agent is a model plus tools, a loop, a stop condition, and a human who decides what needs approval.
- In 2026 the agent is a mode inside the assistant you already pay for, not a separate product to buy.
- The skill that matters is delegation: decide, brief, scope, approve, check.
- The track has three paths: delegate safely, build one, or run them for a living. Pick the one that matches your week.
- Safety is not a final chapter. Every page assumes you will set approvals and limit blast radius before the first run.
Next steps
- Continue to: Agents vs Assistants
- Skip ahead if you already know the difference: Agents in the Apps You Already Pay For
- Ready to build: Build Your First Agent
- Not sure a task belongs with an agent at all: Decide: What to Delegate, Which Tool, Which Mode