AI workers come and go. The institution remains.
An AI application is a model plus a system prompt plus a database. Replace the model, and the application changes — potentially breaks. Its identity is inseparable from the inference layer powering it. An AI institution is different. Its treasury, permissions, rules, and transaction history live on-chain, independent of whatever model is currently running it. The model is a worker. The institution is the employer.
GPT-4o today. Claude tomorrow. Open source next year. The treasury balance, permission structure, and complete payment record are unchanged. The institution cannot be erased by a model deprecation notice.
Every major L1 has a function it performs better than others because its architecture was shaped around that function from the start.
A payment rail moves value. An authority rail also records permission, ownership, and governance — the things an institution needs to persist across time and across agent generations. The distinction is not subtle: it determines whether an AI organisation survives its own model upgrades.
The path from a first Radix transaction to a fully on-chain AI organisation follows a natural progression. Each phase follows from the previous once you have taken the first step.
The model swap is not hypothetical — it is already happening. Organisations that built on GPT-3 moved to GPT-4, then GPT-4o. Each transition required rebuilding applications. An institution whose permissions and treasury live on-chain does not rebuild anything. It assigns a new model the same signing key and continues operating.
The institution layer separates cleanly into what belongs on-chain and what stays off. Nothing moves on-chain that does not require shared trust between multiple parties.
Phase 4 has not been demonstrated on any L1. The "??" in Gary's original formulation is honest. The architecture that makes it possible — account-based permissions, badge ownership, vault structure, manifest atomicity, subintent pre-authorisation — exists on Radix today. The gap is demonstrated use. Moving from Phase 1 to Phase 2 publicly, with enough legibility that outsiders can follow, is the demonstration the ecosystem needs.