AI agents: from experiment to real work tool

Today’s most important AI news

Today’s undisputed protagonist is the rise of AI agents: OpenAI published formal research on how they’re transforming real work, and Google DeepMind launched computer-use capabilities in Gemini 3.5 Flash. In parallel, IBM announced a chip that could keep Moore’s Law alive for another decade, with direct implications for the hardware that runs all these models. It was a day of products and concrete data, not speculation.

AI agents now have numbers: OpenAI shows how they change work

OpenAI published a research paper documenting, with data, how AI agents are changing the nature of professional work. The central conclusion is that agents make it possible to execute longer and more complex tasks than any conversational interaction allowed until now, and that productivity doesn’t grow linearly, but expands into categories of work that previously simply couldn’t be delegated to a machine.

What sets this report apart from the usual marketing material is that OpenAI backs the claims with real usage metrics across different roles: from developers to operations and business functions. The pattern that emerges is consistent: the greatest impact isn’t in doing faster what was already being done, but in tackling work that required coordinating multiple steps and simultaneous information sources.

For a product manager or engineering leader in Latin America, the practical implication is clear: if you’re still evaluating agents as an improved version of the chatbot, you’re looking at the problem through the wrong frame. The relevant question today isn’t “what questions can I ask the AI?” but “which workflows of two or more steps can I redesign so that an agent executes them from start to finish?”

Gemini 3.5 Flash can use your computer, and that’s no small thing

Google DeepMind announced the introduction of “computer use” in Gemini 3.5 Flash, its model designed for high speed and low cost. This means Gemini can now navigate graphical interfaces, click, read screens and execute actions on a real computer, not just generate text describing how to do it.

The choice of model isn’t accidental. Flash is the economy tier of the Gemini family: fast, cheap and built for volume. Bringing computer use to that level of the line suggests Google aims for this capability to be integrated at scale into products and pipelines, not a demo feature reserved for premium models. Anthropic had already shown computer use with Claude last year, and OpenAI has Operator; now Google enters that space more aggressively with a production-oriented model.

For teams building automations or internal flows, this opens concrete options: tasks that today require traditional RPA or fragile scripts could be handled by an agent that simply “sees” the screen and acts. The caveat is that these capabilities still have non-trivial error rates in complex environments, so designing human-in-the-loop supervision remains engineering work that can’t be ignored.

IBM announces the chip that could power the next decade of AI

IBM presented a prototype chip with around 100 billion transistors in an area the size of a fingernail, doubling the density of its previous 2021 technology. The design operates below one nanometer in its critical features, making it one of the most significant advances in semiconductors in recent years.

Moore’s Law — the observation that transistor density doubles roughly every two years — has been declared dead or dying for a decade. This IBM announcement doesn’t resurrect it by decree, but it does show there are concrete technical paths to keep packing more compute into less space and with less energy. The latter matters as much as the former: the energy consumption of the data centers running language models is one of the most cited bottlenecks in the industry.

For those making decisions about AI infrastructure or investment, the message is that hardware that looks limiting today isn’t a permanent ceiling. More capable models require more compute, but the semiconductor industry keeps finding ways to deliver that compute more efficiently. The horizon of what will be possible to run on servers — and eventually at the edge — keeps moving.

In a nutshell

Today’s pattern is one of maturation, not novelty. AI agents have stopped being a research promise and become the centerpiece of the leading labs’ product strategy, with data to back it. Gemini Flash with computer use and IBM with its chip are signals of the same phenomenon seen from two different angles: infrastructure — both software and hardware — is reaching the level needed for agentic automation to work at real scale. What used to be a demo is becoming an engineering and change-management problem.

---

Sources used: https://openai.com/index/how-agents-are-transforming-work, https://deepmind.google/blog/introducing-computer-use-in-gemini-3-5-flash/, https://www.technologyreview.com/2026/06/25/1139696/ibm-unveils-sub1nm-chip/