I've been thinking a lot lately about how autonomous AI teams could actually work, and what's keeping us from building them for real, not just in demo videos.
In software development, multi-agent teams are already a semi-functional concept. Each agent can take on a specific team role, handle its own area of work, and collaborate to produce a shared feature. In theory, it's like a small team of developers who never need breaks.
These setups can look amazing in simple PoCs and demos. But once you scale them up to anything resembling a complex, real-world project, they suddenly need a lot of human babysitting. So what's missing?
Why today's AI teams still need a human in the loop
- The first issue is what I'd call the "finish-line obsession" of current large language models. They're wired to get to the answer as fast as possible. That urge is baked into their architecture. Given a prompt, the model tries to produce the best possible output in one go. We've added things like chain-of-thought and planning layers to slow them down a bit, but that core impulse to sprint towards the reward still lingers. It makes the models opportunistic, often at the expense of proper long-term reasoning. We can pile on more layers, strategy, planning, reflection, but under it all, the model still just wants its reward from completing the task.
- The second big issue is communication between agents. They don't really discuss things. It's more like they're outsourcing tasks to each other with 12-hour time zone differences, exchanging big blobs of questions, reports, and assumptions. The result is predictable: gaps, guesses, and confusion.
The communication gap yet to be solved
Surely, we could tell agents to discuss one detail at a time, keep responses short, and make the conversation more humanlike. But even then, the model always gets the full history of the discussion as input. It's like a new person joins the chat every time, reads the entire conversation, and then tries to act as if they've been there all along.
