Three planes, one spine.

The reasoning loop: memory (owned, profile over episodic, consolidated asynchronously), skills, sub-agent delegation, scheduled jobs, channels (portal, email, Telegram; Teams staged at config), model-agnostic providers with failover.

Registry, policy evaluation, approval routing, append-only audit, spend attribution. Its own application on its own infrastructure: governance survives any single agent.

Public intake (forms + uploads + pipelines), booking, data apps/dashboards. Each is a hardened engine behind one pattern: public inbound → authenticated event callback → guest-lane parse → owner-lane action behind gates.

Agents hold their own service-principal identities; human-triggered work inherits the human's principal; external counterparties are guest-lane by construction — parsed as untrusted content with no authority to steer. Capability is scoped per trigger, not per deployment-wide grant.

Every integration is an MCP server; tools discover at runtime. Adding a system is configuration and a credential, not a rebuild — the pattern proven with our own CRM and task systems (Pulse), Microsoft 365 (native, per-tenant apps), email, and the content engine. If your system speaks MCP, it plugs in.

Each tenant's config, memory, data, and personality isolated (schema-per-agent, row-level security); one codebase upgraded fleet-wide. Deployed in-region — your region, your tenancy, your data. Six deployments run on one stamp today, across independent operating companies of the group.

Provider-agnostic (multiple frontier and open providers) with automatic failover; per-deployment keys make spend attributable per org. Swap models without touching the agent.

Voice channel · agent-to-agent mesh (hub-and-spoke, CoS delegating to specialists) · payments visibility (read-only; payment execution is deliberately out of scope) · enterprise SSO packaging · widened app-builder latitude.