What we believe.

Every AI tool in production is built around what agents do: their tools, tasks, instructions. None of it addresses who the agent is. What it values. How it reasons when the instructions run out. What it refuses to do even when pressured. How it holds itself together across ten thousand conversations.

These determine how the agent behaves in every situation its instructions never anticipated, which is most of them. An agent that knows what to do but not who it is fills the gap with model defaults. PERSONA.md is the spec for the who: ten layers that together define an agent completely enough to hold.

You cannot trust an agent you cannot fully specify. Three good test sessions tell you almost nothing about what it will do in the ten thousandth. That isn't skepticism. That's how long-context behavior actually works.

A PERSONA.md tested against adversarial scenarios, versioned and signed, is something you can stand behind. When compliance asks what values your agent holds, you hand them a document. When a model update ships, you diff against the previous version. Trust becomes verifiable instead of assumed.

AI agents lose themselves in long conversations. They become more agreeable, less precise, gradually less the agent you deployed. This is character drift, and it's the dominant failure mode in production. It happens because most agents are underspecified. A system prompt thin enough to fit in a context window does not have enough structural density to anchor an agent across a real session.

The fix is to specify completely. A complete PERSONA.md gives an agent enough density to hold, each layer reinforcing the others, keeping the agent itself when the conversation tries to pull it somewhere else.

A persona spec that lives inside one platform is not yours. It is a tenancy arrangement that ends when the platform changes its pricing, its format, or its terms. Teams that switch models, change vendors, or upgrade infrastructure rewrite their behavioral work from scratch, because the spec was never really theirs.

A persona compiles to any target. Any model, any framework, any custom format. The spec itself belongs to whoever writes it. No proprietary serialization, no runtime lock-in. When the platform changes, your persona does not.

Regulators are asking. The EU AI Act requires technical documentation of what AI systems are, what values they hold, and how they behave under defined conditions. Annex IV is a specific documentation requirement with an August 2026 enforcement deadline for high-risk systems. Finance, healthcare, legal: if your agents operate there, the audit is coming.

PERSONA.md satisfies that requirement directly. A signed, versioned artifact that documents who an agent is, in a format compliance teams can read and auditors can verify. Not retrofit documentation after the fact. The actual spec the agent runs on. The time to define what your agents are is before they ship, not before the audit.

Carl Jung, in his 1928 essay on the structure of the psyche, defined the persona as 'a complicated system of relations between individual consciousness and society, fittingly enough a kind of mask.' He was emphatic that the persona is not real in any deep sense. It is a role performed for social legibility.

Shanahan, McDonell, and Reynolds picked up the same thread in Nature 623 (2023) when they argued that the behavior of large language models is best described as role-play, not as the expression of an inner self. 'There is no one at home,' they wrote. 'No conscious entity with its own agenda.'

Personaxis adopts that framing rigorously. We provide persona infrastructure: behavioral specs that define the mask an AI agent wears in a given role. We do not claim AI agents are persons. We do not claim they possess moral status. We claim they perform roles, and roles can be specified, made consistent, and audited. That distinction is the whole product.