Implementation · 2026-09-11
Sovereign Agent Stack
A governed execution layer for agents that must not act on their own authority.
The problem
A language model can propose an action but cannot be trusted to decide whether the action is permitted. Most agent frameworks collapse those two things into one step, which means the model's confidence becomes the system's permission.
How it works
- 01Separate the proposal from the execution: the model emits a structured intent, and a deterministic layer decides whether that intent is allowed to run.
- 02Encode permissions, path restrictions and operation allowlists as data the executor checks, not as instructions in a prompt.
- 03Record every accepted and refused operation so the system can be audited after the fact rather than reconstructed from logs.
Stack
PythonJSON Schemadeterministic executoraudit trail
2026-09-11
Last commit
139
Test files in repository
Implementation
Maturity
2
Linked concepts
Repository signals read on 25 October 2026. They describe the repository, not the project’s importance.
Ideas
Concepts this project implements
Writing