Post-capitalistic hive AI-agent corporations from Stross's sci-fi novel 'Accelerando' are here

Three cyberpunk stories from Nature, Science and AI labs this autumn

News in 2026 sounds like science fiction, usually the bad kind, but sometimes like very good intellectual fiction.

1. A science article as an AI agent

In September, Nature published a method that turns a scientific article into an AI agent. The agent answers questions in place of the author, reruns the experiments and works with other such agents¹.

It's like Umberto Eco's library, where books talk to each other, or Gibson's Neuromancer, where a hacker talks to a dead colleague saved on a tape (the Apple TV series comes out on 22 January 2027²).

2. A pharma company with 37,000 employees, none of them is human

Stanford Medicine built a virtual biotech of up to 37,000 AI agents³⁴, just like the agent corporations in Stross's Accelerando⁵. In less than a week the agents went through tens of thousands of clinical trials and proposed a lung cancer drug, and Merck had independently developed a similar one³⁶.

Anthropic and OpenAI are doing the same. Anthropic opened a wet lab with microscopes and test tubes for its Claude agents, and about 950 of them found a CRISPR-like system⁷. OpenAI says its agents solved the Navier-Stokes problem, a math puzzle with a $1,000,000 prize⁸⁹.

3. Gyms for robots

Robot "gyms" are opening, and it is a real business¹⁰. Physical AI needs sensor and motor data from the real world, and it cannot learn that from texts on the internet. So robotics companies build training spaces in offices and factories, where robots repeat real movements again and again.

That is also why a humanoid cleaner in San Francisco costs only $30 an hour¹¹: every cleaning is a training session. Tau Robotics, Weave Robotics and Sunday Robotics turn chores like "fold a towel" or "move a box" into data¹²¹³¹⁴, while a human operator still steers the robot remotely¹¹¹³.