Ask a chatbot how to book a flight and it’ll explain the steps. Ask an agent and it’ll go do it: search fares, compare times, fill in your details, and come back to ask whether to buy.
That’s the shift this group is about. The Agents & Automation group is the 18 orange tiles in the middle of the Periodic Table of AI, and it covers AI that takes action, works through tasks step by step, and sometimes runs without anyone watching.

The last post covered the tools and protocols that connect AI to the outside world. Agents are what happens when AI starts using those tools on its own.
What makes an agent
Agent
An agent is an AI system that works toward a goal by taking actions, checking the results, and deciding what to do next. You give it a job, not just a question. It keeps going until the job is done or it gets stuck and needs you.
Tool Use
Tool use is what lets an agent act at all. Searching the web, reading a file, running code, sending a message: each is a tool. Without tools, an agent is just a chatbot with ambition. This builds directly on function calling from the last post.
Planning
Planning means breaking a big goal into smaller steps before diving in. “Organize a team offsite” turns into finding dates, checking budgets, looking at venues, and sending invites. Good agents also update the plan when something doesn’t work.
Memory
Memory lets an agent hold on to information beyond a single moment. Short-term memory keeps track of what’s happened so far in a task. Long-term memory saves things for later, like your preferences or notes from last week’s work. Since a model’s context window only holds so much, memory systems decide what’s worth keeping.
ReAct Loop
The ReAct loop is the basic rhythm most agents follow: think, act, look at what happened, repeat. The name is short for “reason plus act,” and it comes from a 2022 paper by researchers at Princeton and Google.
Their idea was to let the model alternate between writing out its reasoning and taking actions, like searching for information. The reasoning decides what to look up, and what it finds shapes the next step of reasoning. Grounding the thinking in real lookups cut down on made-up facts compared with reasoning alone.
Agents at work
Computer Use
Computer use lets an AI operate a computer the way a person does. It looks at screenshots of the screen, moves the cursor, clicks buttons, and types. That means it can work with programs that were never built for AI.
Anthropic released computer use in public beta in October 2024 with its Claude 3.5 Sonnet model. At the time, Anthropic called the feature experimental, sometimes clumsy and prone to mistakes, and it encouraged developers to start with low-risk tasks.
Browser Agents
Browser agents are agents that work inside a web browser. They can click through websites, fill in forms, and pull information from pages. They’re handy for chores like comparing prices or gathering details across several sites.
They also carry a particular risk. A web page can contain hidden text written to trick an agent into doing something you didn’t ask for. That’s one reason to keep a close eye on what a browser agent does, especially near logins and payments.
Coding Agents
Coding agents write, run, and fix software. Instead of suggesting one line at a time, they can read a whole project, make changes across many files, run tests, and fix what breaks. They’ve become one of the most widely used kinds of agent, because code gives instant feedback: it either runs or it doesn’t.
Workflow Automation
Workflow automation strings steps together so a process runs on its own. When a form comes in, sort it, pull out the key details, update a spreadsheet, and email the right person. AI makes this kind of automation smarter, because it can handle messy inputs that rigid rules can’t.
Teams of agents
Multi-agent Systems
A multi-agent system uses several agents working together, each with its own job. One might research, another might write, and a third might check the work. Splitting the work this way can help on big or complicated tasks, though more agents also means more ways for things to go wrong.
Subagents
A subagent is a helper agent that a main agent sends off to handle one piece of the job. The main agent might send a subagent to dig through a pile of documents and come back with just the key findings. That keeps the main agent’s workspace uncluttered.
Orchestration
Orchestration is the coordinating. Someone, or something, has to decide which agent does what, in what order, and how their results get combined. Think of it as the conductor in front of the orchestra.
Handoffs
A handoff is when one agent passes a task, along with what it knows so far, to another. A customer service agent might hand a billing question to a billing agent. A good handoff passes along enough context that the next agent doesn’t have to start over.
Skills
Skills are packaged instructions an agent can load when a task calls for them. In Anthropic’s version, a skill is a folder with a main instruction file called SKILL.md, plus any scripts or templates it needs. The agent only reads the full instructions when a task matches, which keeps its memory free for the job at hand.
Anthropic released the Agent Skills format as an open standard in December 2025, and it has since been picked up by other agent tools, including OpenAI’s Codex, Gemini CLI, GitHub Copilot and Cursor. One caution worth knowing: since a skill can include scripts, only install skills from sources you trust.
Running on their own
Triggers
A trigger is an event that starts an agent working. A new email arrives, a file lands in a folder, a customer submits a form. The agent wakes up and handles it, with no one needing to type a request.
Scheduled Tasks
Scheduled tasks run on the clock instead of on an event. Every weekday at 8 a.m., summarize the news in my industry. Every Friday, pull the week’s sales numbers into a report. It’s the alarm clock version of a trigger.
Observability
Observability means being able to see what an agent did and why. Good agent systems keep a record of each step: what the agent was thinking, which tools it used, and what came back. When an agent goes off the rails, that record is how you find out where and fix it.
Autonomy Levels
Autonomy levels describe how much an agent decides on its own. At one end, the AI only suggests and you do everything. In the middle, it acts but asks before anything important. At the far end, it runs start to finish with no check-ins.
The idea echoes how self-driving cars are rated. A standard from SAE International ranks driving automation in six levels, from Level 0, no automation, to Level 5, full automation. AI agents don’t have one official scale like that, but the question is the same: how much do you trust it to drive?
How they fit together
An agent uses tools to act, a plan to stay on track, and memory to keep context. It runs in a ReAct loop: think, act, check, repeat. Bigger jobs get split across subagents and teams, with orchestration and handoffs keeping them in sync. Triggers and schedules let agents start on their own, while observability and autonomy levels keep a person in control of how far they go.
If you take one thing from this post, make it that last part. The more an agent can do, the more it matters to decide what it’s allowed to do without asking.
Where to go next
The next group on the chart is Applications, which covers what people actually use AI for, from translation and summaries to tutoring and customer support. That’s the next post.
For now, click any orange tile and read the full page. Start with the ReAct Loop, since almost every agent you’ll meet runs on some version of it.
1 thought on “AI Agents and Automation: When AI Stops Talking and Starts Doing”
Comments are closed.