There is a question I keep coming back to:
Why doesn't a group of capable agents behave like a capable team?
We have spent the last few years making individual agents dramatically better. They write code, run experiments, browse, plan and use tools. But put several of them on the same problem and something strange happens. They don't add up.
Where the question comes from
I didn't start with agents. I started with people.
Before I worked on AI systems, I spent a lot of time organising humans: building teams, running student societies, starting early ventures. The lesson I took away was that outcomes depended less on how talented each person was, and more on how the group was organised. Who knows what? Who checks whose work? What gets lost in a handover? What happens when the plan stops matching the problem?
Good organisations are not just collections of smart people. They are a kind of intelligence of their own, one that lives in the structure between people. I've come to call this organisational intelligence.
When I moved into AI, I saw the same failures again, only faster:
- Code written by different agents works on its own, then breaks when it is combined.
- Research agents agree on a conclusion without anyone independently checking the evidence.
- Work crosses teams whose agents have different data, tools and permissions, and gets stuck at the boundary.
- One agent fails or hands over unfinished work, and the whole project stalls.
- The problem changes, but the agents keep following the division of labour they were given at the start.
- Everything the group learned about working together disappears when the session ends.
None of these are failures of intelligence. Each agent might be excellent. They are failures of organisation.
The idea
Today, almost every multi-agent system works the same way: a human designs the workflow. Someone decides which agent does what, who passes work to whom, and when to stop. The structure is prescribed.
But that's not how good human organisations get good. They learn. Teams reshuffle roles when something isn't working. People learn whom to trust for what. Experience from one project shapes how the next one is run.
So the question I want to answer is:
Can we turn agent coordination from a workflow we design into a capability the network learns?
Why now
A few years ago this would have been a thought experiment. I think three things have changed:
- Agents can actually do work. Models can now use tools, write and run code, and carry out multi-step tasks, rather than only describe them.
- Agents are becoming connectable. Open interfaces mean agents built by different people, on different models, can increasingly talk to each other, even across organisational boundaries.
- Experiments are getting cheap. Inference costs are falling, and we can now evaluate coordination against working software and completed tasks, not just conversations that sound plausible.
Together, these make it possible to run the experiment for real.
What Systemind is
Systemind is the project I started to test this idea.
The goal is to build adaptive, self-evolving agent networks: networks that form around a task, revise their division of labour and connections as the work changes, and carry verified experience into future work, without ever granting themselves new authority.
We're starting with software engineering and research tasks, because both have clear ways to check whether the work is actually done. We vary who works with whom, how work is delegated and verified, and what experience the network keeps.
The central test is simple to state:
Can an experienced network outperform a fresh one, with the same agents, the same tools, the same permissions and the same compute budget? And does that advantage transfer to new tasks and new collaborators?
If the answer is yes, then organisational intelligence is something we can measure and grow, not just something we describe.
Why it matters to me
If this works, AI organisations become something we can train, not only something we design. A small team could hand a goal to a network that assembles the right expertise, adapts as the work changes, and gets better with experience.
And even if it only partly works, we'll have a map: where learned coordination beats a single agent or a fixed workflow, where it fails, and what it costs. I think that map is worth having either way.
I've believed for a long time that intelligence is not only a property of individuals. It's also a property of how they are organised. Systemind is my attempt to take that belief seriously, and to test it.
If you're thinking about the same question, as a researcher, an engineer, or someone whose work crosses tools and teams, I'd love to talk.