AI Agents: What It Really Means for AI to Go From Chat to Autonomous Work

For years, using artificial intelligence meant one thing: typing into a text box and reading the answer. In 2025 that boundary moved. The major labs shipped systems that don't just reply — they carry out assignments: they browse, open files, run programs and come back with the work done. These are agents, the word that now dominates the industry — used so liberally that a working definition is in order: an agent is software that, given a goal, decides the steps on its own, acts on the world through tools (a browser, a terminal, a calendar) and checks the result of each step before taking the next.

What an agent is (and isn't)

The most useful distinction was published by Anthropic in December 2024, in an engineering guide that has become a reference. It separates workflows — "systems where LLMs and tools are orchestrated through predefined code paths" — from agents: "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." The difference is not cosmetic: in a workflow, a human already chose the path; in an agent, the model chooses it as it goes.

For non-programmers, the most honest analogy is the diligent intern. You ask a search engine a question; you give an intern a complete errand — "put together a comparison of these three suppliers" — along with the tools to do it and clear boundaries. They work alone, but the delicate decisions get signed off with you. No serious office lets the intern sign contracts on day one. Neither do the agents of 2026.

Chatbot, copilot, agent

By that yardstick, the three terms the conversation tends to blur line up as degrees of delegation, not different species. A chatbot answers: text in, text out, and its involvement ends there. A copilot suggests inside your tool while you keep the wheel: it proposes the next line of code or a draft email, and you accept or correct each time. An agent executes: it takes the goal, chains steps — search, compare, write, run — uses external tools and returns a finished result for your review.

Three real agents, with dates

ChatGPT agent (OpenAI, July 17, 2025). "ChatGPT can now do work for you using its own computer," the company announced: a virtual computer with a browser and a terminal, connectable to apps like email and calendars, able to plan a purchase or build a slide deck. The fine print came in the announcement itself: the system asks for explicit permission before consequential actions — a purchase, sending something — and the user can interrupt it or take over at any moment.

Claude Code (Anthropic, February 24, 2025). The company's first "agentic coding tool": it can search and read code, edit files, write and run tests, and push changes to GitHub from the terminal. It launched as a limited research preview, with a design condition Anthropic itself stresses: keeping the developer "in the loop at every step," as "an active collaborator" rather than a replacement.

Copilot coding agent (GitHub, May 19, 2025). Delegation turned into process: you assign it an issue like any other team member and it responds with commits on a draft pull request, running in an isolated environment. The brakes come as standard: its pull requests "require human approval before any CI/CD workflows are run," and the developer who commissions the work cannot be the one to approve it.

The limits, in writing

The limits don't need guessing: they are written down by the very companies selling the agents. Reliability: Anthropic's guide warns that autonomy "means higher costs, and the potential for compounding errors" — a mistake at step two contaminates everything that follows. Supervision: it is not an add-on, it is architecture; all three products above build in explicit permissions, human review, or both, and OpenAI went as far as disabling its agent's memory feature to reduce the risk of prompt injection (malicious content on a webpage that tricks the agent). Cost: agentic systems "trade latency and cost for better task performance," and the official recommendation is to add complexity "only when it demonstrably improves outcomes." Translated: for many tasks, a good chatbot is still the right tool.

The useful takeaway is not futurology but judgment. An agent's maturity today is measured less by what it promises to do alone and more by the quality of its brakes: the three examples in this piece compete on supervision as much as on autonomy. When something is announced as an "agent," three questions clear up almost everything: what tools does it use, what does it check its steps against, and who approves the result. If the answers exist, you have an agent. If not, you have marketing.