Multi-agent collaboration is a governance problem

The demo of multi-agent collaboration always shows agreement. A planner delegates, a researcher retrieves, a critic nods, and the outcome ships. Production looks different. In production, the coverage agent and the risk agent read the same metric store and reach different conclusions, both of which are defensible, neither of which is the CFO's. The institution has simply reproduced its own operating committee inside the fabric, at 3.2 times the decision velocity. That is not a failure of the runtime. It is the runtime working exactly as specified. The question is what happens next, and who has the authority to say so.

Most platforms answer with another agent. A supervisor, an orchestrator, a judge. This is the org chart's oldest reflex, and it produces the org chart's oldest problem: a manager who has never seen the data adjudicating a dispute between two people who have. We tried it. The supervisor agent developed a preference for whichever argument arrived last in the context window, which is recency bias with a service-level objective. Beam search did not help. Monte Carlo tree search made it confident. Self-consistency made it confident more often.

Constitutional AI is the alternative, and we mean it in the boring sense. A constitution is a document that tells agents how to disagree. It names the values that outrank each other, the escalation path when they collide, and the circumstances under which an agent must stop and produce a human. In an institution, that document is called an employee handbook, and the function that maintains it is called HR. We are comfortable with the comparison. Every tenant already has a version of this document. Our job is to make it machine-readable and load-bearing rather than laminated and ignored.

In practice this means every agent in the tenancy is onboarded, not deployed. It receives the constitution, a scope of authority, a list of counterparties it may negotiate with, and a list of counterparties it may only inform. Disagreements are logged as disagreements, with the constitutional clause each party invoked, and PPoE attests that the resolution occurred within the sovereignty boundary. RLHF tunes tone. The constitution tunes precedence. Explainable AI is what makes the difference visible to the change-control board, who will ask.

The objection is that this slows things down. It does, in the places where slow is correct. An agent that can be overruled by a documented principle is an agent whose output can survive an audit, a regulator, and a quarter in which the numbers were wrong. An agent that cannot be overruled is a very fast liability with a Mamba-2.5 backbone, or, on the preview ring, a GPT-9 Universe one. Institutions that have run both tell us the governed version reaches alignment faster, because the argument ends when the constitution says it ends, not when the context window fills.

We think of this as the moment agentic platforms grew up. The first generation asked whether agents could work together. They can. The second generation is asking whether an institution can govern a workforce it cannot see, that operates at machine speed, and that occasionally disagrees with itself in public. That is a governance problem. It has always been a governance problem. The fabric simply removed the option of pretending otherwise.