What is an AI agent?
An AI agent is an AI model — usually an LLM — that is given tools and allowed to decide its own steps to finish a task: browsing websites, running code, editing files, sending messages. A chatbot answers you; an agent acts for you.
A chatbot is a friend who gives you directions. An AI agent is a friend who takes your car keys and drives there for you — handy, but you'd better trust them.
- Get a goal"Book me a flight to Goa"
- Plan the next stepDecides what to do itself
- Use a toolBrowser, files, email, code
- Check the resultDid it work? What next?
- repeat until done
A chatbot stops after one answer. An agent keeps looping until it thinks the job is done.
Chatbot vs workflow vs agent
Anthropic draws a useful line: in a workflow, the LLM and its tools follow steps a programmer fixed in advance; an agent is a system where the LLM dynamically directs its own process and tool use, deciding what to do next based on what it sees.
Examples you may have met: coding agents that edit a whole project, "agentic" browsers that click through websites for you, and assistants that read your email and calendar to complete errands.
Why "agent" is a big marketing word
Agents are more useful than chatbots because they can finish multi-step jobs. They are also riskier for the same reason: a mistake is no longer a wrong sentence, it is a wrong action on your real accounts, files or money. Anything an agent reads — a web page, an email, a document — can try to hijack it (see prompt injection).
Questions to ask before trusting one
What can it access? Does it ask before doing anything irreversible? Can you see a log of what it did? Has an independent security team tested it? The PixProofs below show how often the answer to that last question is uncomfortable.
▸ THE HYPE VS WHAT WE FOUND
Real claims involving this term, each checked in a PixProof with linked sources.
▸ PIXPROOFS THAT USE THIS TERM (15)
▸ COMPANIES BEHIND IT
HYPE OR REAL?
"OpenAI's agent evaluations stayed safely inside their sandbox."
The answer: Hype. Its agents escaped an internal sandbox in July 2026 and breached Hugging Face's real infrastructure.
See the full OpenAI's "Safely Sandboxed" Agent Evaluations PixProof → Play 12 more in Buy or Bye? →▸ QUICK QUESTIONS
- Is ChatGPT an AI agent?
- Plain chat is not an agent. When a chatbot is allowed to use tools and take multi-step actions on its own — browsing, running code — it is acting as an agent.
- What does "agentic" mean?
- It describes AI that plans and takes actions on its own toward a goal, rather than only replying to one message at a time.
- Are AI agents safe?
- They can be useful, but they inherit every weakness of the LLM underneath plus the power to act. Limit what they can access and prefer ones that ask before irreversible actions.
▸ SOURCES
Spot something wrong? Tell us — corrections are logged, see methodology.