What AI agents are, and what has changed since the first ones appeared

A chatbot answers. Ask it for flight options and it returns a list. An AI agent is meant to go further: given a goal, such as booking a trip, it breaks the goal into steps, uses tools to carry them out, checks the results and adjusts. It might read a calendar, compare prices on several websites, fill in a booking form and add a reminder, asking for confirmation at the points that matter.

The Tech Anatomy episode on the subject, published in January 2025, used a kitchen analogy. A chatbot is like a young cook who needs a detailed recipe at every step. An agent is closer to an experienced cook who is told the dish and works out the ingredients, techniques and presentation. The distinction is autonomy: how many decisions the system makes between the human's request and the final result.

Technically, most agents are built from the same large language models as chatbots, wrapped in a loop. The model proposes an action, a piece of software executes it, the result comes back to the model, and the cycle repeats until the task is done or the agent stops to ask a question. What changed in 2024 was the range of actions available.

From research preview to mainstream product

When the episode was recorded, two products defined the category. Anthropic had released "computer use" in public beta in October 2024, letting its Claude model operate a computer by looking at the screen, moving the cursor and typing. The company described it at launch as experimental, "at times cumbersome and error-prone". OpenAI announced Operator on 23 January 2025, a few days before the episode went out, as a research preview limited to its most expensive subscription tier in the United States, able to navigate websites on a user's behalf.

Much has moved since. In July 2025, OpenAI folded Operator into a broader "ChatGPT agent" available inside its main product. A second change was less visible but arguably more important: standard ways for agents to connect to data and software. Anthropic published the Model Context Protocol in November 2024 as an open standard for linking AI assistants to the systems where data lives, and it was widely adopted across the industry during 2025. An agent that can call a hospital scheduling system or a literature database directly does not need to imitate a human clicking through screens.

Healthcare became an explicit target. In January 2026, within days of each other, OpenAI introduced a health space in ChatGPT that lets users connect medical records and wellness apps, and Anthropic announced Claude for Healthcare, aimed at providers, payers and life sciences companies, with connections to medical databases and, for some US users, personal health records.

The use cases the episode described

The episode imagined two scenarios for pharmaceutical teams. In the first, an agent monitors journals, trial registries and regulatory announcements in oncology, and delivers a curated weekly summary to a marketing team. In the second, an agent produces personalised updates for oncologists, in the format each prefers, and adapts its content according to their engagement.

The first scenario can now be built with off-the-shelf tools, and literature monitoring is a task where errors are relatively easy to catch, since a human reads the summary. The second runs into constraints the episode did not discuss. Information sent by a pharmaceutical company to doctors is regulated as promotion in most countries and must be reviewed before it goes out, which sits uneasily with content generated and adapted automatically for each recipient. Under current rules, that points to keeping a human review step for anything that reaches healthcare professionals.

Two readings of the agent wave

One reading sees agents as the point where language models start producing measurable productivity gains. Much work in healthcare and pharma consists of multi-step tasks across several systems: compiling evidence, preparing submissions, reconciling records, scheduling. Agents address exactly that, and the infrastructure built in 2024 and 2025 makes them far easier to deploy than the first browser-driving prototypes.

A second reading is more sceptical. In June 2025, the research firm Gartner predicted that more than 40 percent of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. It also described "agent washing", the relabelling of existing chatbots and automation tools as agents, and estimated that only about 130 of the thousands of vendors claiming agentic products offered the real thing. In a regulated field, every autonomous step is also a point where an error can occur without a human noticing.

What the evidence does not show

Most of what is known about agents in healthcare comes from vendor announcements, demonstrations and benchmarks on simulated tasks. Independent studies measuring time saved, error rates or effects on care in real organisations are still rare. Reliability also drops as tasks get longer, because small errors compound over many steps, and published benchmarks do not always reflect the messy systems found in hospitals.

What to watch

The indicators that would settle the question are independent evaluations of agents in real workflows, with error rates reported and not only success stories; how regulators treat agents that take actions, rather than give advice, in clinical settings; and whether the Gartner forecast of cancelled projects proves accurate when 2027 arrives. The term will probably keep changing meaning in the meantime, which is one more reason to ask, of any product called an agent, which decisions it actually takes on its own.

Listen to the episode

What are AI Agents, the Trend Everyone Talks About?

Listen · 8 min