AI agents are science fiction not yet ready for primetime

It is The stepA weekly newsletter decomposing an essential history of the world of technology. To find out more about everything related to AI, Follow Hayden Field. The stepback arrives in the reception boxes of our subscribers at 8 a.m. HE. Opt for The step here.

It all started with Jarvis yes, that Jarvis by The Marvel Films.

Well, maybe it’s not to start With Iron Man’s assistant from Iron Man, but the fictional system has definitively helped the concept of an AI agent. Whenever I have interviewed people in the AI ​​industry on agency AI, they often indicate Jarvis as an example of the ideal AI tool in many ways – the one who knows what you need before asking, can analyze and find ideas in large data strips, and can offer strategic advice or an execution point on certain aspects of your business. People sometimes disagree on the exact definition of an AI agent, but at the base, it is a step beyond chatbots in that it is a system that can perform complex tasks in several steps in your name without constantly needing back and forth with you. It essentially makes its own list of subtaches which it must complement in order to reach your favorite end goal. This fantasy is closer to being a reality in many ways, but with regard to the real useful user, there are many things that do not work – and may never work.

The term “AI agent” has existed for a long time, but it especially started to tend to technology industry in 2023. It was the year of the concept AI agents; The term was on everyone’s lips when people were trying to overcome the idea and how to make it a reality, but you have not seen many cases of successful use. The following year, 2024, was the year of deployment – People really put the code on the ground and saw what he could do. (The answer, at the time, was not much. And filled with a bunch of error messages.)

I can identify the media threw around AI agents spreading over a specific ad: in February 2024, Klarna, a Fintech company, said that after a month, its AI assistant (fueled by OpenAi technology) had succeeded in the work of 700 full -time customer service agents and automated two -thirds of discussions on the company’s customer service. For months, these statistics appeared in almost all the conversations of the AI ​​industry I had had.

Media threshing has never died and in the following months, each CEO of Big Tech seemed to harass the duration of each call for results. The leaders of Amazon, Meta, Google, Microsoft and a multitude of other companies began to talk about their commitment to create useful and successful AI agents – and tried to put their money where their mouth is to get there.

The vision was that one day, an AI agent could do everything, to book your trip to generate visuals for your commercial presentations. The ideal tool could even, for example, find a good time and a good place to spend time with a pile of your friends who work with all your calendars, food preferences and food restrictions – then reserve the reservation of the dinner and create a calendar event for everyone.

Now let’s talk about the “AI coding” of all this: for years, the coding of the AI ​​has been transportation The agentic AI industry. If you have asked someone on real, successful and non -canceling use cases for AI agents right away And not conceptually in the not too distant future, they indicate the coding of the AI ​​- and it was almost the only concrete thing that they could indicate. Many engineers use AI agents for coding, and they are considered objectively good enough. In fact, in fact, that in Microsoft and Google, up to 30% of the code is now written by AI agents. And for startups like Openai and Anthropic, which burn in cash at high rates, one of their largest income generators is AI coding tools for corporate customers.

So, until recently, the coding of the AI ​​was the main case of real use of the AI ​​agents, but obviously, this does not arise for the everyday consumer. The vision, remember, has always been a kind of AI agent for the “Everyman”. And we are not yet entirely there – but in 2025, we got closer that we have never been before.

Last October, Anthropic launched things by introducing “the use of the computer”, a tool that allowed Claude to use a computer as a human power – navigate, search, access different platforms and perform complex tasks in the name of a user. The general consensus was that the tool was a step forward for technology, but the opinions said that in practice it left a lot to be desired. Quick advance until January 2025, and Openai published the operator, its version of the same thing, and billed it as a tool to fill the forms, order grocery products, reserve trips and create memes. Again, in practice, many users have agreed that the tool was buggy, slow and not always effective. But again, it was an important step. The following month, Openai published Deep Research, an agentic IA tool that could compile long research reports on any subject for a user, and who also made things happen. Some people said the research reports were more impressive than the content, but others were seriously impressed. And then in July, Openai combined the deep research and the operator in a single AI agent product: the chatgpt agent. Was it better than most of the agental tools oriented to consumers who preceded? Absolutely. Was it always difficult to do work successfully in practice? Absolutely.

There is therefore a long way to go to reach this vision of an ideal AI agent, but at the same time, we are technically closer than we have ever been before. This is why technological companies put more and more money in agency AI, by investing in an additional calculation, research and development or talent. Google recently hired the CEO of Winsurf, the co -founder, and some members of the R&D team, in particular to help Google advance its AI agent projects. And businesses like Anthropic and Openai rush to each other on the scale, the scale of the scale, to introduce incremental features to put these agents into the hands of consumers. (Anthropic, for example, has just announced a chrome extension for Claude which allows him to operate in your browser.)

So what happens next is that we will see the Coding of the AI ​​continue to improve (and, unfortunately, to potentially replace the work of many entry -level software engineers). We will also see agent products oriented towards the consumer improve, probably slowly but surely. And we will see agents used more and more for the applications of companies and government, especially since Anthropic, Openai and Xai have all made their debut on AI platforms specific to the government in recent months.

Overall, expect to see more false starts, departures and stops, and mergers and acquisitions while the agent agent competition has resumed (and the continuous media shining bubble). A question that we will all have to ask ourselves for months: what do we really want a conceptual “AI” agent to do for us? Do we want them to replace only logistics or also more personal human aspects of life (that is to say, helping to write a wedding toast or a note for flower delivery)? And how good are they to help with logistics compared to personal stuff? (Answer for the latter: not very good for the moment.)