Bairo GonzalezLeandro · Martinez
Start your guided journey The story

In the world · Artificial intelligence · 17 July 2025

AI agents: the 2025 turning point and what comes in 2026

In July 2025 OpenAI launched ChatGPT agent, which carries out tasks on the computer. In 2026 the AI Index measured the leap in agents. What changes.

bairogonzalez.com team, drawing on Bairo's story · Published

An AI agent is artificial intelligence that stops merely answering and starts doing: it opens websites, fills in forms, builds spreadsheets and asks for permission before acting. On July 17, 2025, OpenAI launched ChatGPT agent, which brings a browser, a terminal and app connections together inside ChatGPT. In April 2026, Stanford University's AI Index recorded that agents' success on real-world tasks rose from about 20% to 77.3% in one year.

What happened

The path was fast. In January 2025, OpenAI introduced Operator, an agent that uses the browser the way a person would. In July, it merged Operator and deep research into a single product, ChatGPT agent. According to coverage by MacRumors and TechCrunch, it:

  • researches and writes reports, builds presentations and spreadsheets;
  • browses websites, filters results and handles login screens;
  • connects to services such as Gmail and GitHub;
  • was designed to ask permission before any consequential action, and the person can interrupt the task at any moment.

OpenAI itself acknowledged new risks. It classified the model as high-capability in biology and chemistry, turned on extra monitoring and disabled memory in agent mode to reduce the risk of leaks through malicious instructions hidden in web pages.

Other companies moved in the same direction throughout 2025, with agents in the browser, in the office and in programming. In early 2026, the AI Index 2026 measured the effect: agents' success rate on real-world tasks jumped to 77.3%. The same report points to the other side, the unevenness of capability: household robots still get only 12% of tasks right.

Why it matters

A chatbot hands back text and the person does the rest. An agent does part of the work for them. That changes what a small business can do without hiring: check prices, put together a quote, organize documents, reply to a supplier.

It also changes the size of the possible mistake. When AI only wrote, the error stayed on paper. When it acts, the error can become a purchase, a sent email or a wrongly filled form. That is why the rule of asking permission before acting matters as much as the agent's capability.

In Bairo's view

For Bairo Leandro Gonzalez Martinez, agents are the most concrete form of the idea he defends: AI is not a chatbot, it is an extension of you. In his reading, an agent that acts with the permission of whoever decides extends a person's reach without taking the wheel away from them.

Bairo also says he started without knowing how to code and spent six months at home building with the tools of his own time. For him, agents shorten that path for those who come next: the distance between having an idea and seeing it work has become shorter.

His bet, however, is not on one agent that does everything. It is on many specialist agents, coordinated. That is the design of CLAIN, the orchestrator he founded: some research, others check, one brings together what survived and a person decides. The goal of 142 niche AIs in 142 days, announced on October 1, 2026, is written as a goal.

Where this meets the ecosystem

Bairo's ventures use, or propose to use, cells of agents, always with a human at the end:

  • At CEASA Bank, a financial platform for agribusiness that operates on the infrastructure of an authorized institution, the proposal is an agent that checks reference prices and prepares the quote and invoice for a load within minutes, for the trader to review and decide.
  • At SOSPROCESSOS, under development, agents organize rulings and scenarios, and one of them tries to knock down the legal argument before the lawyer takes it forward.
  • At Hospital Triage, also under development, pre-consultation is done with a clinical assistant and validated by a health professional.

The common thread is the same one OpenAI wrote into its own product: the agent asks permission before what matters.

Sources

Read also

More on this topic

In the world: Artificial intelligence