Core Concepts

Tribunus is built on four foundational concepts. Understanding these will help you make the most of the platform, whether you are running a single agent on your laptop or orchestrating a fleet of inference workers.

Compile-time architecture

Unlike traditional AI frameworks that schedule operators at runtime, Tribunus Compute resolves the entire execution graph — kernel selection, memory layout, operator fusion — at compile time. The result is a self-contained binary that has no runtime branching for scheduling decisions. This eliminates jitter, reduces memory overhead, and makes inference behavior fully deterministic for a given input shape and model.

The compile-time approach also enables a rigorous testing model: the same compilation pipeline that runs during development runs in production. Any regression in precision or performance is caught before deployment.

Numerical oracle

The numerical oracle is a critical component of Tribunus’s correctness guarantee. It is a reference implementation of every supported operator, computed at high precision (float64 or higher). During compilation, the oracle validates that the optimized kernel output matches the reference output within a configurable tolerance.

For models with security or regulatory requirements, the oracle can be configured to enforce strict bounds on numerical deviation. This makes Tribunus suitable for domains where approximate inference is insufficient — financial analysis, scientific computing, and regulated industries.

Agent system

Tribunus Desktop provides a multi-agent runtime where each agent is an independent session with its own model, tool set, and orchestration loop. Agents can communicate, delegate subtasks, and share context through structured channels.

The agent system is extensible via plugins. A plugin can add new tools (code analysis, web search, database queries), new model providers, or custom orchestration strategies. The plugin API is language-agnostic: plugins communicate with Desktop over a JSON-RPC protocol.

Plugin architecture

Plugins are the primary extension mechanism. Each plugin is a self-contained process that Desktop discovers and manages. Desktop ships with a catalog of official plugins — Git integration, file system tools, MCP support — and you can write your own in any language that speaks JSON-RPC.

Plugins are sandboxed: each runs in its own process with declared capabilities. Desktop enforces the capability boundary and displays a grant prompt when a plugin requests an operation outside its declared scope.