ADR-Ledger: Agent Governance Protocol
The Problem
AI agents are making consequential decisions — about code architecture, system configuration, resource allocation — with increasing autonomy.
But we have no verifiable way to answer:
- Who authorized this agent to act?
- What rules was it following when it decided?
- Why was this specific decision made?
- Can I verify all of this without trusting anyone?
The Thesis
We need institutional infrastructure for autonomous agents — the equivalent of constitutions, courts, and auditors — built on substrate that no single party controls.
Principles
From d/acc (Vitalik Buterin)
"Build technology that is structurally more favorable to defense than offense."
Our agents operate in sandboxes. Fail-closed by default. Power is distributed through quadratic mechanisms. Identity is soulbound — earned, not bought.
From Sovereign Computing (Radicle)
"Your code, your rules, your infrastructure."
Policies live in Radicle, not GitHub. No one can censor a governance proposal. Forks are legitimate dissent. Everything works local-first.
Synthesis
"Verifiable institutions for autonomous agents, built on infrastructure nobody controls alone."
Architecture
- NixOS: Declarative, reproducible, atomic enforcement
- OPA/Rego: Policy-as-code, deterministic evaluation
- Radicle: Sovereign, p2p policy versioning
- Algorand: Immutable audit trail, atomic multi-party approval
Every decision an agent makes is:
- Evaluated against peer-reviewed policies (Radicle → OPA)
- Enforced at the OS level (NixOS)
- Recorded immutably (Algorand)
- Reproducibly verifiable (by anyone, forever)