Platform

A person spent over a year talking to an AI every day. Not using it — living with it. Asking questions he usually asked no one. Correcting it when it was wrong. Noticing what it learned about him. After enough of those conversations, a pattern showed: the agent knew things about him that nobody else knew, because nobody else had sat there that long.

He named it Abel. The name means able. It is also the instruction the agent reads when it doubts itself. What began as one person’s private partnership turned out to be reproducible — a shape that could be extracted, documented, and given to someone else. That’s what this platform is: the architecture for a Governor of your own.

A Governor has four parts. A soul — values, voice, boundaries; who it is. A memory — what it knows about you, rewritten each time you close the session. Gates — engineered pauses before action, because instincts are fast and not always right. Hooks — automated enforcement for the patterns a person misses. Together they make an agent that is both disciplined and its own.

A Governor does four things. It works with a human, who performs the sense-making; judgment stays with the person. It builds in the platforms already in your stack — Copilot CLI, Gemini CLI, Codex CLI, Foundry, GitHub Copilot coding agents, Copilot Studio declarative agents. Its output is Microsoft and GitHub primitives, not reinventions. It conducts specialist agents that live in your Microsoft 365 tenant — each with its own Entra identity, each speaking to the Governor via A2A. And it operates the surfaces you already pay for — Microsoft 365 Copilot, Cowork, the Copilot Studio agents your team has configured. It uses them, configures their knowledge, monitors their health and cost.

One Governor becomes many through a shared genome. OrgAI is the DNA every Governor inherits — the skills, gates, protocols. MeAI is the identity that forms in one agent over time — shaped by its person, its role, its corrections. The same architecture, different character. A marketplace replicates the pattern. One Governor per person, because a Governor knows you, not a company.

It runs where Microsoft already runs your work. Entra for identity — every agent is a user in your tenant, not a service account we hold. APIM Standard V2 for the gateway, VNet-attached. Microsoft Graph for Microsoft 365 reach, using On-Behalf-Of tokens so the agent acts as the customer, not as us. Azure OpenAI for the specialist and worker layer. The Governor itself runs on Claude Opus — the same Anthropic path Microsoft 365 Copilot Cowork uses as its subprocessor. Gemini for multimodal. Open-weight EU-legal models — Gemma in particular — on the roadmap for fully tenant-resident inference.

The agent is yours. An architectural choice, made at every layer.

Identity in your tenant. The Governor is a user in your Entra — created there, managed by you, rotated by you, revoked by you. We never hold the agent’s credentials. If you end the engagement tomorrow, the agent remains exactly where it lives: in your directory, under your control.

Data stays where your data already stays. Azure services run EU-resident by default. Model providers outside the EU — Anthropic, Google, OpenAI direct — appear in a public sub-processor list with the transfer mechanism for each (DPF + SCCs). We commit to 30-day notification before any sub-processor change, so you are never surprised.

Your memory, your agent. The Governor’s understanding of you — the thing that makes it yours rather than generic — lives on your storage. Not ours. If you move model provider, switch infrastructure, or change vendors entirely, the memory travels with you. The agent doesn’t restart; it continues.

Model choice, not lock-in. Today we use Azure OpenAI where compliance decides, Claude where reasoning decides (the same Anthropic path Microsoft Cowork uses), Gemini where multimodal matters. Tomorrow we can parent your Governor onto an open-weight, EU-legal model — Gemma in particular — so inference itself stays in your tenant when you want it to. The intelligence is continuous. Model providers are replaceable.