AI has made individuals much faster, but a team of 10x individuals is not a 10x team. The opposite can be true: the team gets slower because the processes get clogged. I think this is the problem to work on now. It’s why we need to push for more multiplayer AI that is built for teams and not only for individuals.
I hit this personally. At Giskard, I led R&D on AI safety and security, with a team of AI researchers and engineers who understood AI much better than the typical engineering team. We adopted AI tools early, and individual output went up fast. Yet, initially, team output did not follow.
One person could open 40 pull requests in a day, and nobody had time to review them. People started side projects without checking with product or business. Everyone used different tools in different ways. Nobody – well, but me – was doing anything wrong. Processes were not designed for coordinating at that level of speed and capabilities.
Speed needs continuous steering
Research on expertise shows that hours of practice alone don’t make you an expert. What counts is deliberate practice, which means practicing with a specific goal and paying attention to feedback.
Teams need something similar, which I call deliberate velocity. Forty pull requests in forty directions are noise. A weekly sync used to be enough to keep a team aligned. With AI, three people can build three incompatible versions of a feature in an afternoon. Coordination is more important than before.
Disparity in adoption is a problem
In many cases, the group is as fast as its slowest member. Making a small number of individuals capable of moving faster can even be detrimental to the team’s velocity, as it generates friction and frustration that may damage the ability to work together.
The problem is that tools like Codex or Claude Code are built for one technical power user. They introduce concepts like agents, sub-agents, worktrees, skills. Grok Bot introduces armies of specialized agents. Engineers generally like that, but most people in a company do not.
A business rarely wins on technique. It wins on domain expertise, which sits with the people who know the customers and the business. Today most of that knowledge stays outside the AI loop: to bring those people in, the interface has to be simple and natural.
Individual AI tools are only a stepping stone towards AI for teams. Value comes from keeping everyone in the loop.
Security can’t be only gates
Our processes were designed for human speed, with big reviews and approval gates at the end. When an agent does in an hour what used to take a week, the review queue becomes the bottleneck.
I’m not arguing for less control. Coming from AI security, I’d argue for more. Agents that act for a team need clear permissions. Review has to happen during the work, in small steps, be done by the right person at the right time, and be automated when the stakes are small. Ideally nobody notices it.