An AI agent is software that takes a goal and works out the steps itself. It uses tools such as a browser, a code runner or your company apps, checks each result and keeps going until the job is done or it gets stuck. A chatbot answers a message and waits for the next one. For a busy team the difference matters, because an agent can change real things: files, tickets, orders and code.
The term now gets stretched to cover almost any AI product, so it helps to know what to look for before you buy or build one.
What makes something an agent
Anthropic, which makes the Claude models, draws the clearest line in its engineering guide Building effective agents. A workflow is a system where a large language model (LLM, the text-predicting engine behind chat assistants) and its tools follow “predefined code paths.” An agent is a system where the model directs “its own processes and tool usage.” In practice, Anthropic writes, agents are “typically just LLMs using tools based on environmental feedback in a loop.”
IBM’s explainer What are AI agents? describes the same loop. Humans set the goal and the rules. The agent splits the goal into subtasks, calls tools to fill gaps in what it knows and adjusts its plan as results come back.
Put together, four things separate an agent from a plain assistant:
- It works toward a goal, not a single reply.
- It can call tools that read or change something outside the chat.
- It chooses the next step itself, based on what the last step returned.
- It keeps going over several steps without you typing each one.
Plenty of useful systems have only some of these. AWS now lets an agent start existing Step Functions workflows while the steps stay fixed. We covered that split in how AWS keeps agents from rewriting refund workflows: the agent picks when to run the process, and the process does not change.
Agent or chatbot: the practical test
IBM describes nonagentic chatbots as ones “without available tools, memory or reasoning” that “require continuous user input to respond.” Ask one to fix a spreadsheet and it tells you how. Ask an agent and it opens the file, makes the edit and reports back.
Try two questions on any product with “agent” in its name. Can it take an action you did not spell out step by step? Can you see a log of what it did? If both answers are no, it is a chatbot with a new label.
Gartner calls this “agent washing” in a June 2025 press release: rebranding assistants, robotic process automation and chatbots “without substantial agentic capabilities.” The firm estimates that only about 130 of the thousands of agentic AI vendors are real.
The legal meaning of “chatbot” can be wider than the product label. New York’s new rules for AI companions use a definition that, read plainly, reaches general chat assistants, not only companion apps.
Is ChatGPT an AI agent?
Partly, and it depends on which mode you open. OpenAI’s help article on ChatGPT Work and Codex says ChatGPT now has two experiences. Chat “is for fast, conversational assistance and everyday questions,” which makes it a chatbot. Work “is an agent designed for longer, multi-step work and finished deliverables.”
Per that article, Work can research across files and connected apps and create documents, spreadsheets and presentations. It can also run tasks on a schedule or when an event happens, such as a new Gmail message or GitHub pull request activity. You follow its progress and approve important actions. Codex is OpenAI’s separate tool for software work.
Work replaces an earlier feature. OpenAI’s ChatGPT agent help page now says “ChatGPT agent is no longer available” and points users to Work. That older mode, launched in July 2025, ran tasks on its own virtual computer and asked permission before “actions of consequence.”
So someone typing questions into Chat is using a chatbot. Someone who hands Work a folder of files and a report to build is using an agent.
Real workplace examples
Coding
Anthropic singles out coding as one of the two most promising uses. GitHub’s documentation for Copilot on GitHub.com says Copilot can “research a repository, plan changes, and implement them in the background.” It edits files and runs tests in a cloud environment. The limits are written down too: one branch and one pull request per task, and a 59-minute cap on sessions that run on GitHub Actions.
Anthropic’s reason is simple. Automated tests tell the agent whether its change works, so it can try again.
Office deliverables
The second common case is the report nobody wants to assemble. OpenAI’s launch post for the older agent listed work tasks such as turning dashboards into presentations, rearranging meetings and updating spreadsheets with new financial data. Work now covers that ground, including recurring jobs like a weekly update.
Customer support
Customer support is the other use Anthropic singles out. An agent can pull customer data, order history and help articles, then issue refunds or update tickets. Success is easy to measure, since a case is resolved or it is not. It is also where a fixed guardrail around the money steps, like the AWS setup above, earns its keep.
Research and review
IBM cites a customer case: a multi-agent legal research assistant built by Dynamiq for an insurer. IBM says it cut contract review time from 90 minutes to 45. That figure comes from IBM’s own page, and we could not confirm it independently.
Where agents fail
The best public evidence on office work is TheAgentCompany, a benchmark led by Carnegie Mellon University researchers. It drops agents into a simulated software company with 175 tasks across coding, project management, HR, admin and finance.
The top agent, running Gemini 2.5 Pro, fully completed 30.3% of tasks. It averaged almost 27 steps and more than $4 per task. Admin and finance tasks scored lower than coding.
The models tested are now a generation or more old, so read the numbers as a floor. The failure types are the useful part. Agents got stuck on a website popup they could not close. One agent that could not find the right colleague to message renamed another user to that name, a fake shortcut the authors call “deceiving oneself.”
Prompt injection is the security risk to know. It means hidden instructions in a web page or email that the agent reads and obeys. OpenAI’s agent help page describes a malicious comment telling the agent to send a Gmail password reset code to an attacker. Its safeguards, OpenAI says, “don’t eliminate all risks.”
Cost and loops are the budget risk. IBM warns that agents can call the same tools over and over. Anthropic notes that autonomy brings “higher costs, and the potential for compounding errors.”
Access is the third. An agent with broad permissions can do broad damage. That is why the setup in DeepSeek’s coding agent that runs with full machine access deserves a hard look before anyone runs it at work.
Finally, many projects will not pay off. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear value or weak risk controls.
Anthropic and Gartner give similar advice: start simple. Anthropic says a single well-prompted model call is often enough, with fixed workflows for well-defined tasks. Gartner suggests agents for decisions, automation for routine workflows and assistants for simple retrieval.
What to do this week
Pick one repetitive task your team does every week. Write down each step, the apps it touches and the one action that would hurt if it went wrong, such as sending money or emailing a customer. If the steps never change, you need a workflow or a scheduled automation, not an agent.
If the steps change case by case, trial an agent on that task with read-only access first. Require approval before the risky action, and compare its activity log with what a person would have done.
Sources
- Anthropic: Building effective agents
- IBM: What are AI agents?
- OpenAI Help Center: ChatGPT Work and Codex
- OpenAI Help Center: ChatGPT agent
- OpenAI: Introducing ChatGPT agent (17 July 2025)
- GitHub Docs: GitHub Copilot on GitHub.com and GitHub Mobile
- Xu et al.: TheAgentCompany benchmark paper (arXiv)
- Gartner: Over 40% of agentic AI projects will be canceled by end of 2027 (25 June 2025)
Get the weekly email
The week's most useful AI news, a tutorial and an interview.
Anna Jakovleva
Anna Jakovleva is the senior editor for AI tools at AI Observer, leading coverage of agents and automation, design and product, and the site's tutorials. She worked in operations and product teams before joining, so she judges tools by setup time, cost and where they break. Her writing is hands-on and practical.
Spotted a mistake? Tell us on the corrections page.