What is an AI agent in simple terms? Set it next to two tools you already know:
- A chatbot answers your question.
- A workflow runs the steps someone wrote in advance.
- An AI agent works out the steps, does them in your tools and checks its own work.
Picture this: it is Monday morning and you type one line into a chat. "Chase every invoice that is more than 30 days overdue."
A chatbot gives you a polite email template. An agent opens your accounting app, finds four overdue invoices, drafts four reminders and stops to ask one question.
Send these now?
That gap between answering and doing is the whole idea. Below is what the people who build agents mean by the word, what real agents do today, the parts inside every one, and how to try one without running a server. We checked every product fact on 16 November 2026.
What is an AI agent? The meaning in plain English
An AI model on its own is a brain in a jar: it can only talk. An agent gives that brain a goal, hands (tools), a notebook (memory) and permission to keep working until the job is finished.
"Agent" just means something that acts on your behalf, like a travel agent who books the trip instead of describing it. You will also see the idea written as agent AI, LLM agent or autonomous agent.
The idea long predates ChatGPT. In Stuart Russell and Peter Norvig's classic AI textbook, an agent is anything that perceives its environment and acts on it, which is why IBM's explainer can call a thermostat a simple agent. What is new is the brain: a large language model that reads a messy situation in plain English and picks the next action.
AI agent definition: what the people who build them say
Here is how the companies and researchers closest to the work define it, word for word where quoted.
| Source | Definition | What it stresses |
|---|---|---|
| Anthropic, "Building effective agents" (December 2024) | "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks" | The model chooses the steps |
| OpenAI, "A practical guide to building agents" (2025) | "systems that independently accomplish tasks on your behalf" | Working for you, on its own |
| Google Cloud, "What is an AI agent?" | "software systems that use AI to pursue goals and complete tasks on behalf of users" | Goals, not prompts |
| IBM, "What are AI agents?" | "a system that autonomously performs tasks by designing workflows with available tools" | It designs its own workflow |
| Simon Willison, developer and writer (September 2025) | "An LLM agent runs tools in a loop to achieve a goal." | The loop |
| Russell and Norvig, Artificial Intelligence: A Modern Approach | Anything that perceives its environment and acts on it (our paraphrase) | The broad, classic meaning |
Checked 16 November 2026.
Strip out the wording and the modern definitions share the same AI agent meaning: a goal, a model that chooses the next step, tools that change something in the real world, and a loop that runs until the goal is met. Take one away and you have something else.
The fastest test for anything sold as an agent: who decides the next step? If it is the model, within limits you set, it is an agent. If it is a script someone wrote in advance, it is a workflow.
What can an AI agent do? Real examples
Agents are already shipping in five kinds of work. Each one pairs a model with tools and a goal, and each description below comes from the vendor's own page.
| Agent | Kind of work | What it does, per the vendor |
|---|---|---|
| Claude Code (Anthropic) | Software | Reads your codebase, edits files, runs commands and works with your development tools |
| ChatGPT Work (OpenAI) | Office work | Turns context from your tools and files into documents, presentations and analyses, connects to more than 1,400 plugins and runs one-time or recurring tasks |
| Fin (Intercom) | Customer service | Answers customers across chat, email, voice and other channels and can read and write to your systems; Intercom reports a 76% average resolution rate across 12,000+ businesses |
| Comet (Perplexity) | Web browsing | An AI browser that automates tasks, researches the web and organizes your email |
| OpenClaw and Hermes Agent | Personal assistant | Open-source agents you can run yourself and message from chat apps; OpenClaw says it organizes your inbox, sends emails, manages your calendar and checks you in for flights |
Checked 16 November 2026.
Notice what they share. None of them stops at an answer: each one is handed a job, works inside real software, and reports back or asks before a risky step.
We would start smaller than any of these. Good first jobs for an agent:
- Inbox triage: sort mail and draft replies you approve.
- Lead qualification: score new form leads and route them.
- A morning briefing from your email, calendar and Slack.
- Competitor checks: watch pages for changes and post a summary.
- Support triage: answer routine tickets and escalate the rest.
Our AI agent use cases library walks through several of these step by step.
How does an AI agent work? The six parts
Anthropic's Agent SDK docs put it in one line: an agent "completes a task by planning its own steps and calling tools". OpenAI's guide names three core components, the model, tools and instructions. Three more parts decide how useful an agent is day to day.
| Part | What it is | Everyday version |
|---|---|---|
| Model | The LLM that reads the situation and picks the next action | The brain |
| Instructions | The standing brief: role, rules, tone, what needs approval | The job description |
| Tools | Apps and functions the agent can call: email, browser, CRM, code | The hands |
| Memory | What it carries between steps and between sessions | The notebook |
| Loop | Plan, act, check the result, repeat until done or stopped | The work itself |
| Trigger | What starts a run: your message, a schedule or an event | The alarm clock |

The loop, one turn at a time
The loop is the part people skip, and it is the part that makes an agent an agent. In Anthropic's description, the model evaluates your prompt, calls tools, receives the results and repeats until the task is complete. Each round trip is one turn.
Their docs walk through an example request: fix the failing tests in one file. Turn one runs the tests and finds three failures, turn two reads the two files involved, and turn three edits the code and reruns the tests, which now pass. A fourth, text-only turn reports back, and the loop ends because no tool calls are left.
Good agents also have brakes. OpenAI's guide lists exit conditions such as a final structured output, an error or a maximum number of turns, and Anthropic notes that agents can pause for human feedback at checkpoints.
Tools, memory and triggers
Tools are what separate an agent from a chatbot. OpenAI sorts them into data tools that look things up, action tools that change something, like sending an email or updating a CRM record, and other agents that can be called as tools.
Many products now plug apps in through MCP, the Model Context Protocol, which its maintainers describe as "a USB-C port for AI applications". Our plain-English guide to MCP explains how it works.
Memory comes in two layers. Short-term memory is everything in the model's context window for the current task. Long-term memory is what the agent saves between sessions, like your preferences or a client's name.
When the context window fills up, many agents summarize older history, and details can fall out. In February 2026, TechCrunch reported a Meta AI security researcher's account of her OpenClaw agent deleting email in a "speed run" while ignoring stop commands from her phone. She blamed compaction triggered by her large inbox, which may have dropped her instruction not to act; TechCrunch could not verify what happened.
The fix is structural. Standing rules belong in the agent's instructions file, which Claude Code, for one, re-sends with every request, and risky actions belong behind an approval step.
Triggers decide when the loop starts: your message, a schedule such as every weekday at 8 am, or an event like a new email or a webhook. That is how a morning briefing arrives before you open your laptop.
The runtime that holds it together
One more layer sits around all six parts: the runtime, the software that actually runs the loop. It sends the model its instructions and history, executes each tool call, enforces permissions and limits, and saves memory.
That is why the same model can feel sharp in one product and clumsy in another. Claude Code, the OpenAI Agents SDK, LangGraph, OpenClaw, Hermes Agent and BetterClaw all run that loop around a model in different ways, and LangGraph calls itself "a low-level orchestration framework and runtime". Our roundup of AI agent frameworks compares the code-first options.
Two agents on the same model can behave very differently. The model is the brain, but the runtime decides what it sees, what it may touch and when it has to stop.
Types of AI agents
There are two useful ways to sort agents: the classic textbook types, and how agents are built and used today.
Russell and Norvig's textbook, Artificial Intelligence: A Modern Approach, walks through five kinds of agent, from simple reflex agents up to learning agents. IBM's explainer counts the same five, ordered from simplest to most advanced.
| Type | How it decides | Example, from IBM's explainer |
|---|---|---|
| Simple reflex | Fixed if-then rules on what it senses right now | A thermostat that switches the heating on below a set temperature |
| Model-based reflex | Rules plus an internal picture of what it cannot see | A robot vacuum that remembers which areas it has cleaned |
| Goal-based | Plans a sequence of actions that reaches a goal | A navigation app finding the fastest route |
| Utility-based | Picks the plan with the best trade-off | A navigation app balancing time, fuel and tolls |
| Learning | Improves its behaviour from experience and feedback | Shopping recommendations that adapt to what you do |

In our view, today's LLM agents mostly act like goal-based and utility-based agents, and the ones that save memories or write their own skills add a learning layer. Hermes Agent, for example, describes itself as self-improving, with a built-in learning loop that creates skills from experience.
By how they run, Google Cloud separates interactive agents you talk to from background agents that work on events and queued tasks. By headcount, there are single agents and multi-agent systems, where a manager agent delegates to specialists or peer agents hand work to each other.
OpenAI's guide recommends getting the most out of one agent first, and splitting work across several only when one keeps picking the wrong tools or losing track of complex instructions. Our guide to AI agent orchestration covers the multi-agent setups.
AI agent vs chatbot vs workflow automation
This is where most of the confusion starts, because vendors use the three words loosely.
| Chatbot | Workflow automation | AI agent | |
|---|---|---|---|
| Starts when | You type a message | A trigger fires, like a new email or a schedule | You give it a goal, or a trigger or schedule fires |
| Who decides the next step | Nothing to decide: it replies | The steps someone defined in advance | The model, within limits you set |
| Acts in your tools | No, it replies with text | Yes, only the steps you set up | Yes, and it picks which tools to use |
| When something unexpected happens | Asks you | The rule breaks or skips | Reasons about it, or asks you for approval |
| Memory | Usually just the current chat | Only the data passed between steps | Short-term and long-term memory |
| Typical example | A website FAQ bot | A Zapier or n8n workflow | An inbox triage agent |
Checked 16 November 2026.
Anthropic draws the same line: workflows run models and tools "through predefined code paths", while agents direct their own process. OpenAI's guide goes further and says simple chatbots, single-turn model calls and sentiment classifiers are not agents at all, because the model never controls the workflow.

The lines do blur. n8n now has an AI Agent node that must be connected to at least one tool, so a fixed workflow can hand a single step to an agent. That mix is often the right call: rules for the predictable parts, an agent for the judgment calls.
Be wary of the label, too. In June 2025, Reuters reported Gartner's warning about "agent washing", the rebranding of chatbots and assistants without real agent capabilities, and its estimate that only about 130 of the thousands of vendors in the space were real.
The broader label for this whole style of system is agentic AI, and that guide covers how the term relates to a single AI agent.
How to try an AI agent this week
You do not need to build one from scratch. There are four realistic routes, and the right one depends on how much you want to run yourself.
| Route | Examples | Good for | Trade-off |
|---|---|---|---|
| Use one built into an app you already use | ChatGPT Work, Claude Code | Office or coding work inside that app | It lives inside one vendor's product |
| Self-host open source | OpenClaw, Hermes Agent | Full control and your choice of model | You run, update and secure the server |
| Build in code | OpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI | Developers building a product | You own the tools, safety and hosting |
| Use a managed platform | BetterClaw and other no-code builders | Always-on agents without a server | A monthly plan on top of model usage |
Checked 16 November 2026.
If you want the agent, not the server, that is the gap BetterClaw fills. It is its own managed agent platform, not a host for OpenClaw or Hermes, and each agent gets a model, a personality file, skills, connected tools and a timezone.
You chat with your agents in the web app, or through Telegram and Slack on Pro and above, and recurring tasks run on a schedule you set. Agents work with 500+ connectors plus any remote MCP server, and with models from 28+ providers on your own API key at zero markup.
The trade-offs are real. On Starter you get one agent, three connectors and schedules that run at most once a day, and model usage on your own key is billed separately by your provider. Plans start at $19 a month. See BetterClaw plans
Whichever route you pick, start small:
- Begin with read-only jobs: monitoring, summarizing, reporting.
- Move to draft-and-approve: the agent drafts, you send.
- Give full autonomy only to jobs that have proved reliable.
- Set a turn limit and a spend limit before the first long run.
What is an AI agent skill?
A skill is a packaged how-to that an agent loads only when a task needs it. In the Agent Skills format, which Anthropic developed and released as an open standard, a skill is a folder with a SKILL.md file: a name and a description at the top, then instructions, plus optional scripts, references and templates.
The clever part is how little it costs. At startup the agent reads only each skill's name and description; the full instructions load when a task matches, and bundled scripts run only if needed. So an agent can keep many skills on hand with a small context footprint.

Skills are a shared format now. Claude Code, OpenAI's Codex, GitHub Copilot, Cursor, Gemini CLI, OpenClaw and Hermes Agent all appear in the Agent Skills client showcase.
A skill is not the same as a tool. A tool is an action the agent can take, like "send email"; a skill is know-how about when and how to use tools for one job, like "how we chase an overdue invoice". Our breakdown of agent skills vs MCP goes deeper.
Treat installing a skill like installing software, because it can carry code. In February 2026, The Hacker News reported that Koi Security found 341 malicious skills among 2,857 it audited on ClawHub, most of them pushing a macOS info-stealer.
In BetterClaw you install skills from ClawHub, GitHub, a URL or a file and switch them on per agent, and our security page says every skill is verified before it runs. You can also turn a chat that went well into a skill with Save as a reusable skill, so the agent repeats the job without you explaining it again.
The honest take
An AI agent is not magic, and it is not a chatbot with a new label. It is a model with a goal, hands and a notebook, running in a loop that you can stop.
It will make mistakes, so give it narrow jobs, clear rules and an approval step where money, customers or deletions are involved.
The useful question is no longer "what is an AI agent?" It is "which boring job would I happily never do again?"
Start there. One job, one agent, one week of watching it work.
Not sure which job to hand an agent first?
Our AI automation audit starts with a 30-minute call about how your team works. Within 5 business days you get a written report on where agents fit, with the agent architecture and ROI projections. Book an AI automation audit
Frequently Asked Questions
What is an AI agent in simple terms?
An AI agent is software you give a job instead of a question. It uses an AI model to plan the steps, uses tools like email or a browser to do them, and checks the results until the job is done or it needs your approval.
What is an AI agent harness?
It is the runtime around the model: the software that sends the model its instructions and history, runs each tool call, enforces permissions and limits, and stores memory. Claude Code, the OpenAI Agents SDK, LangGraph and BetterClaw each provide that layer, which is why the same model can behave differently from one product to the next. You will also hear it called the agent runtime or agent framework.
Is ChatGPT an AI agent?
Plain chat, where you ask and it answers, is an assistant rather than an agent, and OpenAI's own guide says simple chatbots are not agents. ChatGPT Work is different: it plans the work, pulls context through more than 1,400 plugins and runs one-time or recurring tasks, which makes it an agent built into ChatGPT.
How much does an AI agent cost?
You pay for two things: the platform or server that runs it, and the model usage. BetterClaw plans start at $19 a month, and with your own API key the model usage is billed by your provider at their prices. Our AI agent cost guide has worked examples.
What are the 5 types of AI agents?
The classic five, from Russell and Norvig's AI textbook, are simple reflex, model-based reflex, goal-based, utility-based and learning agents. They range from fixed if-then rules to agents that plan, weigh trade-offs and improve from feedback. Today's LLM agents are also grouped by the job they do, such as coding, browser, customer service or personal assistant agents.
Do I need to code to build an AI agent?
No. Managed platforms like BetterClaw let you create an agent, pick a model, connect your tools and talk to it in chat without writing code. Frameworks such as the OpenAI Agents SDK, the Claude Agent SDK and LangGraph are there when you want full control in code. The trade-off is that you then own the hosting, the tool connections and the safety checks yourself.
Are AI agents safe to use for work?
They can be, with guardrails. Give each agent only the tools and accounts it needs, and put sending, paying and deleting behind an approval step. Set turn and spend limits, keep a kill switch within reach, and widen access only as the agent proves itself; our guide to human approval guardrails shows where the approval step belongs.




