Splendor

Overview

A public guide to Splendor, the governed runtime for autonomous agent loops.

Splendor is a kernel-grade runtime substrate for governed autonomous agent loops. It gives agents explicit identities, versioned state, append-only traces, verified action boundaries, quotas, messages, replay, and governance hooks while running on ordinary Unix-like systems.

Splendor does not replace your operating system, model stack, planner, or product UI. It sits underneath them as the enforcement and audit layer for persistent autonomy.

Runtime model

Percepts -> Policy -> Constraints -> Gateway -> Verifiers -> Adapter -> Outcome -> State Commit -> Trace

Policies and SDK code propose work. The Rust runtime enforces identity, permissions, quotas, verifier decisions, state commits, trace emission, and safe replay behavior.

What Splendor standardizes

  • Identity: tenant, agent, run, tick, action, state, trace, message, work order, approval, fleet, node, and instance IDs stay distinct.
  • State: every meaningful runtime state change is committed as a state graph node or explicit state reference.
  • Trace: decisions, denials, outcomes, messages, approvals, state commits, and replay facts are recorded as runtime contract data, not informal logs.
  • Actions: side effects are mediated by the Action Gateway and verifier chain before adapters execute.
  • Replay: replay is inspect-only by default and does not repeat filesystem, network, database, shell, webhook, artifact, or device side effects.
  • Coordination: typed messages and scoped delegation allow multi-agent work without implicit permission inheritance.

What Splendor is not

  • Not a bare-metal operating system.
  • Not a chat-first agent framework.
  • Not a replacement for product control planes or enterprise admin UI.
  • Not a hard real-time robot, drone, PLC, or motor controller.
  • Not a universal distributed memory or consensus system.

Documentation map

On this page