What agents can do today

So what do AI agents actually do in 2026? Here are the main categories where they genuinely work today, with example tools. (Specific product names change fast — focus on the capabilities.)

• Coding agents — read a codebase, plan a fix, write the code, run tests, and even open a pull request. Examples: Claude Code, OpenAI's coding agents, Cursor, GitHub Copilot's agent, Devin, Replit Agent. These are among the most mature and useful agents today.
• Computer-use / browser agents — "see" a screen or web page and click, type, and navigate like a person to complete tasks in ordinary software. Examples include Anthropic's Claude computer use / Claude in Chrome, OpenAI's agent mode in ChatGPT, and Google's Gemini computer use.
• Research agents ("deep research") — spend several minutes autonomously browsing many sources and produce a structured, cited report. Available in ChatGPT, Gemini, and Perplexity.
• Workflow / automation agents — carry out multi-step business processes across connected apps (CRM, ticketing, finance). Examples: Salesforce Agentforce, Microsoft Copilot Studio, and others. This is where a lot of business interest is.
• Customer-service agents — handle routine support chats, resolve common cases from a knowledge base, and escalate to a human when unsure.

What they have in common: each takes a goal, uses tools in a loop (last lesson), and produces a result with less step-by-step hand-holding than a chatbot needs. The best current uses are narrow and well-bounded — a specific coding task, a research report, a defined workflow — not "run my whole life."

A reality note (expanded next lesson): a lot of what's marketed as an "agent" is really a helpful assistant that retrieves and drafts but doesn't truly act on its own. Genuinely autonomous, acting agents are real but still best in bounded tasks with a human watching. So as you hear "agent" everywhere, ask: does this actually take actions, or just answer?

The mindset: AI agents genuinely work today in several categories — coding, computer/browser use, deep research, workflow automation, and customer support — where they pursue a bounded goal using tools, with less hand-holding than a chatbot. The capabilities are real and useful, especially for narrow, well-defined tasks. But keep a critical eye: much of what's labeled an "agent" is really an assistant, and even true agents work best in bounded tasks with oversight — which the next lesson makes clear.