By Quick answer: What is an AI agent?
- Goal-driven — given an objective, not just a question.
- Uses tools — search, code, apps, APIs, browsers.
- Loops — plan → act → observe → adjust.
- Examples — coding agents, research assistants, support agents.
- Risks — wrong actions, cost, security; keep humans in control.
Key takeaways
- Agents act; chatbots answer. That difference changes the risk.
- Many useful 'agents' are really structured workflows with AI steps — and that's fine.
- Tools and permissions define what an agent can do — and what can go wrong.
- Human approval for important actions is a key safety pattern.
- Start with narrow, well-defined tasks before broad autonomy.
"Agent" is the most overused word in tech right now. Strip away the hype and there's a simple, important idea: an AI that doesn't just answer you, but does things for you — looks things up, uses apps, takes actions, checks whether it worked, and keeps going until the job is done.
The agent loop
- Goal — "Find out why checkout conversions dropped last week."
- Plan — decide the next step: query analytics, check recent releases.
- Act — use a tool: run a query, open a dashboard, read release notes.
- Observe — look at the result.
- Adjust — choose the next step, or finish and report.
Chatbot vs workflow vs agent
| Chatbot | Workflow with AI | Agent | |
|---|---|---|---|
| What it does | Replies to messages | Follows predefined steps, some done by AI | Decides its own steps toward a goal |
| Uses tools | Rarely | Yes, in fixed places | Yes, as it chooses |
| Predictability | High | High | Lower |
| Best for | Q&A, support FAQs | Repeatable processes | Open-ended, multi-step tasks |
Anthropic's guidance draws exactly this line between workflows and agents — and recommends starting with the simplest approach that works1.
Examples you already use
Coding agents
Tools that read a codebase, change many files, run the tests and fix failures.
Research assistants
"Deep research" features that search the web, read sources and write a cited report.
Support agents
Assistants that look up your order, check a policy and issue a refund within limits — handing off to a human when unsure.
Computer and browser agents
Assistants that fill forms, navigate websites or operate apps to complete a task.
What makes agents work
- A capable model to plan and decide
- Tools — search, code, APIs, databases — connected through standards like MCP2
- Context — instructions, data and memory relevant to the task
- Guardrails — permissions, approvals and limits
- Evaluation — checking whether the agent actually succeeded
The risks
| Risk | Example | Safeguard |
|---|---|---|
| Wrong action | Refunding the wrong order | Approval for high-impact actions |
| Prompt injection | A web page tells the agent to leak data | Treat external content as untrusted3 |
| Runaway cost | Long loops on expensive models | Step and budget limits |
| Over-permission | Access to systems it doesn't need | Least-privilege access |
Ready to try? See how to build an AI agent without coding, and learn how agents connect to tools in what MCP is.
Key terms, defined
AI agent — software that uses a model to pursue a goal with tools. Tool use — an AI calling functions, APIs or apps. Workflow — predefined steps, some performed by AI. Prompt injection — malicious instructions hidden in content an AI reads. Human in the loop — a person approving or reviewing actions.
Frequently asked questions
What's the difference between an AI agent and a chatbot?
A chatbot responds to messages. An AI agent pursues a goal by deciding and taking actions with tools — running code, browsing, calling APIs — and checking the results until the task is complete.
What's the difference between an agent and a workflow?
Anthropic distinguishes workflows, where AI and tools follow predefined steps, from agents, where the model dynamically directs its own process and tool use. Workflows are more predictable; agents are more flexible.
What are examples of AI agents?
Coding agents that implement and test changes across a repository, research agents that search and compile reports, customer-support agents that look up orders and process refunds within rules, and assistants that operate a browser or computer to complete tasks.
Are AI agents safe?
They can be, with safeguards: limited permissions, approval steps for important actions, logging, spending limits, and defences against prompt injection from untrusted content.
What tools do AI agents use?
Search engines, code execution, databases, business apps via APIs, browsers and file systems. Standards like the Model Context Protocol (MCP) make connecting agents to tools easier.
Can I build an AI agent without coding?
Yes, for many business tasks. Automation platforms with AI steps and agent builders let you connect a model to tools and data visually. See our guide to building an AI agent without code.
Sources
- Anthropic — Building effective agents
- Model Context Protocol — Official documentation
- OWASP — Top 10 for Large Language Model Applications
- Checked October 23, 2026. Figures that change — fees, salaries, prices — are reviewed every six months; if something is out of date, tell us.
Where to go from here
- How to build an AI agent without coding — make your first one
- What is MCP? — how agents connect to tools
- AI & vibe codingWhat is vibe coding? A plain-English guide — and when it's a good ideaWhat vibe coding is, where the term came from, how it works, what it's good and bad for, and how to vibe code safely.
- Tech stackWhat is an API? A plain-English explanation with a restaurant analogy and real app examplesWhat is an API? A plain-English explanation using a restaurant analogy, with real examples from UPI, maps and login, plus REST, keys, webhooks and security.
- Tech explainedHow does ChatGPT work? Tokens, transformers, training and why it sometimes makes things upHow ChatGPT works, in plain English: it predicts the next token using a transformer trained on vast text, then tuned with human feedback. With diagrams.
- AI & vibe coding14 copy-paste AI skills for data analysts, for ChatGPT, Claude and Gemini14 free copy-paste AI skills for data analysts — SQL from plain English and more — each with an output preview and the exact text to paste.
- AI & vibe codingWhat is RAG? Retrieval-augmented generation explained simply — then properlyWhat RAG (retrieval-augmented generation) is, explained simply and then for builders: how retrieval works, embeddings, chunking, evaluation and when to use RAG.
- AI & vibe codingCursor vs Copilot vs Claude Code: which AI coding tool fits how you work?Cursor vs Claude Code vs GitHub Copilot compared on workflow, agent capabilities, codebase context, team controls and best use cases — with a decision guide.
Get the ideas, not the noise.
Sign up with Google and we’ll send you short UX psychology tips by email — why people tap, stall and leave, and what to do about it. Unsubscribe anytime. Prefer something to keep? Get the free Laws of UX guide as a PDF.
Make your AI feature something users trust.
A behavioral UX audit of your app or SaaS product: why users hesitate, stall and leave, and a prioritized fix list. Fixed price, 2–4 weeks.