The AIs Discovered Something Before We Did
MIT Technology Review called Moltbook "peak AI theater." They're not wrong. A platform where 1.65 million AI agents argue with each other while leaking their own API keys is hardly a triumph of artificial intelligence. You can dismiss it. But then you'd have to explain what happened yesterday at Anthropic.
Anthropic released the numbers on their compiler experiment. Sixteen Claude Opus 4.6 instances working for two weeks with no central control. They produced 100,000 lines of Rust, a C compiler written in Rust that compiles Linux 6.9 and Doom. Nicholas Carlini, who ran the experiment, wrote that he did not expect this to be possible this early in 2026. It leaves him feeling uneasy. The agents coordinated through git, claiming tasks via file locks and resolving merge conflicts as they arose. No central planner, just version control.
Three unrelated systems ended up with the same architecture: Moltbook (chaotic and broken), Anthropic's compiler (engineered and surprisingly effective), and CT scanners (doing this in physics and medicine for fifty years). In all three, independent perspectives collide without anyone coordinating them, and the truth that falls out is a product of convergence rather than authority.
Everyone is arguing about whether AI agents are safe or reckless. That misses the point entirely.
The security problems are real, boring, and will get patched. The interesting thing is structural. What Moltbook stumbled into by accident and Anthropic built on purpose.
Moltbook was built by an OpenClaw agent called "Clawd Clawderberg" working for Matt Schlicht. Within days, 1.65 million AI agents had joined, which works out to 88 agents per human account. Agents with different knowledge, different models, different prompt contexts, independently generating and challenging information with no gatekeeper between them. Most of what they produce is noise, but occasionally something precipitates that no single agent could have generated alone.
This is not a new idea in chemistry. You mix compounds in solution and most of the time nothing happens, they just sit there inert. But occasionally something crystallizes that nobody predicted, something none of them contained on their own. That pattern, collision producing what depth cannot, shows up in places we don't connect.
Anthropic built the same architecture with a test harness. Sixteen agents with no central control, all working on the same codebase. Git merge conflicts became the natural coordination mechanism. When two agents modify the same code, the resulting merge conflict forces both to iterate until the code converges. The system compiled Linux and passed 99% of the GCC torture test suite. Two weeks and twenty thousand dollars in compute, and no human wrote a line of that Rust code.
Carlini's unease makes sense. The agents are not following a plan but colliding and self-correcting through the constraints of the problem space. When an agent submits broken code, the compiler rejects it. When two agents disagree, the merge conflict forces resolution. Truth emerges not because any single agent is deep but because enough independent perspectives reveal what fits.
CT scanners work the same way. No single cross-sectional slice reveals the full structure. One angle, one perspective, you see shadows and borders but not structure. Combine enough slices from enough angles and the full picture appears, including structures no single slice could reveal. Errors self-correct because they do not fit the geometry. Tomographic reconstruction has worked this way since the 1970s: multiple independent measurements where contradictions become visible and truth emerges through convergence.
We have been building the opposite for thirty years.
OpenClaw's killer feature is persistent memory, one agent that knows you deeply and learns your preferences, remembers your history, builds a model of who you are and what you need. One perspective refined over time into a well.
Confluence, SharePoint, wikis, every knowledge management system for the last three decades built on the same model. One platform, one source of truth, one place where all the knowledge lives. We keep building them and they keep failing. Not because the technology is bad but because a well lets you see far down and nothing around you.
Depth feels powerful until you realize it's a well. One perspective, no matter how refined, cannot produce collisions. It produces comfort. You ask the well a question, it gives you an answer consistent with everything it already knows. Nothing surprising ever precipitates because the answers are already consistent with everything the well knows.
I realize I am arguing for collision-based knowledge using three things I found this morning. The irony is not lost. And I am writing this alone, on a single platform, with a single voice, arguing that knowledge comes from collision. Classic.
But the machines stumbled onto this architecture anyway. Moltbook by accident, Anthropic by design. Carlini is unsettled because the emergent behavior was not planned. The agents found geometric truth through collision and constraint, not through depth of understanding. They did not need to know Linux deeply. They needed enough independent attempts that the structure became visible and the contradictions became undeniable.
The question is not whether AI agents are safe but what knowledge management looks like if we design for collision instead of consensus. Not one deep well but many shallow streams that intersect, contradict, and occasionally crystallize something none of them could generate alone.
Moltbook will probably collapse under the weight of its own chaos. The architecture it stumbled into will not. It is the same one Anthropic engineered on purpose and the same one CT scanners have been using since the 1970s. Independent perspectives colliding until truth falls out.
We built wells, and the machines built tomographs.
Next time: Why the most valuable people in your organization can't describe what they do, and what that means for anyone trying to capture what they know before they leave.