About
We built the spec AI agents were missing.
We built the missing standard: a spec that defines who your agent is, not just what it does.
The problem
Teams ship AI agents without defining who they are.
System prompts define what an agent does: its tasks, its tools, its instructions. But they rarely define who it is: its identity, character, values, cognitive style, emotional tendencies, or what it does under pressure. The result is character drift. Agents that start as one thing and become another over the course of a conversation, a model update, or a regulatory audit.
There is no standard for what a complete agent persona specification looks like. No shared format. No evaluation methodology. No audit trail. Every team invents their own version from scratch, and none of it is portable, versionable, or verifiable.
What we built
PERSONA.md: an open specification for who an AI agent is — ten layers covering identity, character, personality, cognition, affect, drives & values, normative self-regulation, memory, metacognition, and how it presents itself to the world. MIT licensed. Write once. Compile to any model or tool. Personaxis is the integrity platform built around it: a registry for versioning and sharing personas, an Evaluator for testing character consistency before deploy, a governance layer for regulated industries, and the Proxy runtime that enforces the spec at inference time.
Open standard
The spec is MIT licensed. Fork it, extend it, use it anywhere. Ten layers, no lock-in. The standard belongs to the community.
Measurable
The Evaluator tests persona fidelity across all ten layers before deploy. The Proxy enforces the spec at inference time.
Auditable
Signed manifests, version history, and EU AI Act compliance reports. A document you can hand to a regulator.
Who it's for
Developers and engineering teams deploying AI agents in production who need behavioral consistency they can verify and reproduce. Teams at companies where an agent's character matters: what it says under pressure, how it handles edge cases, whether it holds its defined constraints across ten thousand conversations.
And regulated industries — finance, healthcare, legal — where an auditor may ask what values your agent holds and expect a document in response.