AI agents are redefining the fundamental unit of knowledge work, shifting from single interactions to tasks that can be delegated and autonomously executed over extended periods. According to an OpenAI report, the Codex tool now accounts for 99.8% of the total tokens generated weekly within companies, with non-developer usage increasing 137-fold. Google DeepMind’s Co-Scientist multi-agent system, published in Nature, has already produced tangible scientific breakthroughs in areas such as liver fibrosis treatment and cellular aging research, with the potential to significantly shorten the cycle of scientific discovery.

Article author and source: 36Kr

The AI agent is redefining the fundamental unit of knowledge work—from single interactions to tasks that can be delegated and autonomously executed over extended periods.

This is the first sentence of a recent report released by OpenAI. The report states that agents can operate independently for hours, calling tools, interacting with external environments, and continuously iterating until the task is completed. This leap in capability is driving agents to become the most powerful AI tools in the workplace.

Similarly, Google DeepMind recently unveiled its latest Co-Scientist study, published in Nature—a multi-agent system built on Gemini, in which multiple specialized agents collaborate to simulate the full cycle of scientific reasoning. The study documents several real-world cases demonstrating Co-Scientist’s tangible impact across multiple scientific disciplines.

AI agents are not just making human work faster; they are simultaneously transforming how humans work and quietly expanding the boundaries of what human work can achieve.

How has the way of working been changed?

This report by OpenAI documents the internal adoption trajectory of the AI agent tool Codex, using itself as the case study.

In the first few months after Codex was released to the public, ChatGPT remained the default AI tool for employees even within OpenAI. Until August 2025, less than 10% of token usage by regular employees was on Codex. But this situation rapidly reversed thereafter.

By 2026, Codex had become the primary AI tool across every department at OpenAI—not just for engineers, but also for legal, finance, and recruiting teams. It currently accounts for 99.8% of the company’s weekly output in tokens. OpenAI believes this trend reflects the future of work, as agent tools become more capable and accessible, making such adoption the new norm.

On one hand, task durations continue to increase. The report estimates the "human hours" corresponding to Codex requests. By May 2026, 80.6% of users had submitted tasks to Codex requiring more than 30 minutes of human work, 70.2% had submitted tasks requiring more than one hour of human work, and tasks requiring over eight hours of work showed the fastest growth rate.

Initially, Codex was primarily used to quickly answer questions and generate code snippets; now, users are beginning to delegate entire tasks to it—research, analysis, and workflow setup. OpenAI notes that as Codex’s ability to handle long contexts and independent tasks continues to improve, user behavior is quietly shifting: from brief, immediate interactions toward more complex, longer-term task delegation.

On the other hand, the potential boundaries of work are being expanded as a result. It is not surprising that programmers were among the first to adopt Codex—it is, after all, a tool built around programming. However, reports show that since August 2025, usage among non-developer individual users has increased 137-fold, and among non-developer enterprise users, it has grown 189-fold. Additionally, in other non-programming technical roles, more than a quarter of the content generated by Codex users falls into engineering or programming-related tasks. This means that tasks previously requiring support from technical teams—such as automation, data processing, tool building, and debugging—can now be handled directly by non-programmer employees through Agent.

These changes have direct relevance for companies redesigning workflows, employees assessing which skills are more valuable, and researchers understanding how AI is reshaping the labor market. OpenAI states that when people can seamlessly use powerful agent tools, they will naturally apply them to longer, more complex tasks that span multiple functional boundaries. Over time, this is likely to become the shape of future work.

The scientist's deputy can also be an Agent.

If OpenAI’s report illustrates a shift in how knowledge work is delivered in the workplace, Google DeepMind’s Co-Scientist study published in Nature demonstrates that AI agents are playing a substantial role across various scientific research tasks.

Google DeepMind states that Co-Scientist aims to solve the "needle in a haystack" problem of identifying the correct scientific hypothesis amid vast amounts of information, because "every major scientific breakthrough often begins with a correct hypothesis." Co-Scientist is a multi-agent system built on Gemini, composed of specialized agents that collaborate to simulate the complete cycle of scientific thinking—generating hypotheses, critically reviewing them, and iteratively evolving them. The system operates in three stages: generation, debate, and evolution. First, agents propose initial hypotheses and cluster them for diversity; then, "virtual peer reviewers" critically evaluate these hypotheses; finally, the top-ranked directions are continuously refined to produce research proposals for scientists to review. The entire system is coordinated by a "supervisory agent" that breaks down high-level research goals into actionable steps and drives multiple agents to explore in parallel.

The most distinctive feature of Co-Scientist is its method of validating hypotheses. The system draws inspiration from the competition mechanisms of AlphaGo and AlphaStar—but instead of having AIs play chess or video games, it enables agents to engage in scientific debates. All candidate hypotheses are entered into a "creativity tournament," where they are continuously filtered, eliminated, and evolved through pairwise comparisons and simulated debates, while deeply cross-referencing scientific literature and professional databases to ensure that every remaining hypothesis is logically sound and empirically supported. The majority of computational resources are dedicated to this validation phase.

The report documents multiple real-world cases demonstrating the tangible impact Co-Scientist has already made in scientific research. For instance, one scientist and their team used an AI agent to accelerate the exploration of treatments for liver fibrosis, uncovering previously overlooked drugs; another team reduced the time required to analyze massive screening datasets from months to just days in research on reversing cellular aging; and a company in the field of aging biology generated new hypotheses based on Co-Scientist, which were later validated in laboratory experiments.

Scientists involved in the research said that AI agents offer various benefits in scientific work, but the most important is their ability to enhance efficiency, significantly shortening the time required to achieve scientific breakthroughs.

Google DeepMind stated on their official blog that their AI agent is "designed to be a research partner, not a substitute for scientific or clinical expertise." Meanwhile, OpenAI emphasized in their report that AI agents are not just about "speeding up" work, but about expanding the scope of tasks each person can achieve. Both focus on how humans can collaborate with AI agents to accomplish more complex work.

But the boundaries of collaboration will inevitably be continuously redrawn. When agents can take on tasks that previously required specialized skills, how should workflows be redesigned? As functional boundaries begin to blur, which abilities will become more valuable, and which will be reassessed? When the speed of generating scientific hypotheses increases by orders of magnitude, which fields will be the first to see breakthroughs?

The real question is where people should devote their time and energy.

Source: Sequoia Capital Insights