Architecture

RARS is a stateful, neurosymbolic execution engine that turns your codified business into a programmable runtime. This page summarizes the core architectural concepts. Each section links to detailed documentation.

The Operating Model

Your organization's knowledge, its domains, rules, operations, and integrations, is codified into formal specifications called matrices. A matrix is a complete, versioned operating model for a specific business domain: HR, finance, procurement, project management, or anything else your organization does.

Matrices are composable. When you install them into a workspace, they merge into what we call the global matrix: a monolithic "code repository" that RARS has access to.

RARS doesn't see HR and finance as separate tools. It sees one connected operating model where a project's budget comes from the finance matrix, its team members come from HR, and its deliverables are tracked in project management.

This composition follows a holarchical structure: each matrix is both a complete, self-contained whole and a part of a larger whole. Your workspace becomes a global operating model assembled from composable parts, and your organizational architecture adapts to the model rather than the other way around.

The Runtime

RARS is not a collection of autonomous agents passing messages between each other. It is one composed agent, constructed dynamically from the matrices installed in your workspace. Matrices are activated just-in-time: a context doesn't load your entire operating model upfront. As the work requires new domains, RARS activates the relevant matrices on demand, loading their logic and integrating them into its reasoning.

Your codified operating model exposes components that we can load into a prompt, creating a new purpose-built agent for what you're working on at each moment.

Execution is powered by a NeuroSymbolic AI engine with two integrated layers:

• The probabilistic layer (LLM reasoning): understands natural language, makes judgment calls, plans multi-step approaches, and adapts when things change.
• The symbolic layer (Your Operating System): runs deterministic workflows against your operating model, dispatches actions based on type hierarchies, enforces validation, and tracks provenance.

These layers are interleaved within the same execution loop. RARS reasons, plans, programs, and executes. The results feed back into context, and it reasons again. The AI doesn't call tools or react step-by-step. It writes and executes programs of interdependent stages against your codified business model.

Shared state, not message passing

Every actor in a context, whether human, AI, or external system, operates on the same knowledge graph through a collaborative runtime. Alignment across domains isn't coordinated through middleware or meetings. It is structural.

Persistent memory

RARS maintains a dual-context architecture that separates the LLM's prompt context from the persistent knowledge graph. The graph serves as long-term memory that the AI can query at any point, giving it gigabytes worth of context without being limited by the token window.

Data sovereignty

Your context data lives in your context. Poliglot's serverless isolation model ensures workloads are isolated at every level: your data doesn't leak to other contexts, workspaces, or tenants.

Trust and Security

AI operating systems break traditional access control assumptions. A single user request can trigger a chain of operations spanning multiple domains, multiple agent identities, and multiple systems of record. RARS addresses this with three architectural decisions:

The Identity Model: every actor in the system, human users, RARS itself, and matrix-provided agents, has a formal identity. Delegation chains track exactly who initiated what and through which path.

Situational Access Control: permissions aren't static role assignments. They evaluate the live state of the business at the moment of execution. Not just "does this role have permission," but "does the current business situation support this action."

Contractual AI: every action has an explicit input/output contract. AI output is validated against the same schema as a direct API response. Combined with deterministic plan execution, this replaces the step-by-step ReAct loop with a model where the AI plans a complete workflow and the symbolic engine executes it with formal guarantees.