Skip to content
VURT

VURT/Runtime infrastructure

Infrastructure for intelligent systems.

VURT builds runtime infrastructure for state, memory, permissions, routing, and controlled execution across AI systems.

See the architectureEngineering in the open↗
Section through a system running on VURTModels and agents enter through a context and memory band, pass into the VURT runtime holding persistent state, and act on tools, devices, and the surrounding environment.MODELS / AGENTSCONTEXT + MEMORYVURT RUNTIMESTATETOOLS / SOFTWARE / DEVICES / ENVIRONMENT
00Thesis

Intelligence alone is not a system.

As models take on longer-running, higher-consequence work, the hard problems move underneath the model. A capable system still has to know what is true, remember what happened, hold to what it is allowed to do, and leave behind a record someone can check.

Those are runtime concerns. Prompts do not enforce them and context windows do not preserve them. They belong to a layer that outlives any single call.

Governed by the runtime

  • 01what they know
  • 02what persists
  • 03what they may access
  • 04which model or tool executes
  • 05what actions are permitted
  • 06what state changes
  • 07what happened
  • 08what can be inspected or controlled
01Infrastructure

Six concerns, one runtime.

Each answers a question a system must be able to answer about itself. They are separable, and they are meant to compose.

  • What is true right now?

    A typed, durable record of the system and the environment it operates in — held outside any single model call, so it survives restarts, model swaps, and long-running work.

02Architecture

A control layer between the model and the world.

Models and agents do not act on the environment directly. They act through a runtime that holds state, resolves authority, routes work, applies effects, and records the result.
ModelsAgentsContext + memoryPERSISTEDVURT RUNTIMEState01Permissions02Routing03Execution04Telemetry05ToolsSoftwareDevicesEnvironmentEFFECTS APPLIED THROUGH THE RUNTIME

Runtime modules

Holds what is true.

Every actor operates against a shared, typed representation of system and world state. State is not reconstructed per call; it is the thing calls operate on.

03Engineering principles

How we decide what the runtime should do.

These are engineering commitments, not claims about outcomes. They are the constraints VURT designs under.
  • 01

    Explicit state over hidden state

    If a system depends on it, it should be represented, typed, and readable — not implied by a prompt or reconstructed from a transcript.

  • 02

    Controlled execution over unconstrained autonomy

    Capability is granted, not assumed. The runtime decides what an actor may do; the actor decides what to attempt.

  • 03

    Inspectability over opacity

    Any decision that changes state should leave a record precise enough to reconstruct it later.

  • 04

    Deterministic boundaries around probabilistic systems

    Models are probabilistic. The interfaces around them — permissions, transitions, effects — should not be.

  • 05

    Model independence

    State, memory, and authority outlive the model that used them. No layer should assume a single provider or generation.

  • 06

    Composable infrastructure

    Each concern stands alone and composes with the rest. Adopting one part should not require adopting all of it.

04Open development

Engineering in the open.

VURT builds foundational control infrastructure for intelligent systems: the runtime layer beneath increasingly autonomous AI. Work in progress is published where it can be read, run, and argued with.

github.com/vurtai

The VURT GitHub organization — runtime components, interfaces, and technical work as it is developed.

Open the organization