The Gap in Most AI Deployments
Most AI deployments have no laws. They have system prompts. They have guidelines. They have good intentions baked into the initial configuration.
But system prompts drift. Guidelines get overridden. Good intentions do not survive contact with edge cases at 3am when no one is watching.
The Sovereignty Protocol takes a different position: if you want your AI systems to behave consistently — across every interaction, every model update, every new agent you deploy — you need constitutional laws. Not suggestions. Not reminders. Laws.
What Agent Laws Are
Agent Laws in the Sovereignty Protocol are structured YAML definitions that live in the governance core of the platform. They define the rules that every agent in your instance operates under, with four distinct categories:
Constitutional Laws
The highest-order rules. These define the fundamental identity and purpose of the agent system. Constitutional laws are not about specific behaviours — they are about what the system is and what it exists to do. Changing a constitutional law is an auditable, deliberate act that requires explicit operator approval.
Ethical Laws
Rules about how agents interact with people, data, and sensitive contexts. Ethical laws define things like: how the system handles personally identifiable information, what it refuses to do regardless of instruction, and how it escalates situations it cannot resolve within its defined authority.
Technical Laws
Operational rules about how agents execute tasks. These cover things like: which external services are permitted, what output formats are acceptable, how errors are handled, and what constitutes a valid result. Technical laws prevent agents from taking shortcuts that produce correct-looking but incorrect outputs.
Operational Laws
Rules about the agent's behaviour within the platform itself: logging requirements, notification thresholds, escalation paths, and the conditions under which a task should be paused for human review rather than completed autonomously.
The GEMINI.md Constitution
At the heart of the Sovereignty Protocol's governance model is a file called GEMINI.md. This is the constitutional document — a human-readable record of the foundational laws that govern the entire system.
GEMINI.md is not a configuration file in the traditional sense. It is a living document that defines the character of the system: its purpose, its limits, its values, and its obligations. Agents that operate within the Sovereignty Protocol are trained to treat GEMINI.md as authoritative — it is the source of truth for what the system stands for.
The name is deliberate. Like a constitutional founding document, it is meant to be read, understood, and referenced — not just processed.
Why Immutability Matters
The critical word in the law definition is immutable. Not "default." Not "recommended." Immutable.
An AI agent that can talk itself out of its own governance rules is not governed. An agent that can be prompted into ignoring its constitutional laws by a sufficiently clever instruction is not safe.
The Sovereignty Protocol enforces laws at the infrastructure level, not just the prompt level. Laws are not part of the conversation context — they are checked before context is assembled. An agent cannot be instructed to ignore them because the instruction arrives after the law check has already run.
This is a fundamental architectural choice. It means the governance layer is not a feature of the model — it is a constraint on the model's operating environment.
Self-Assessment Loops
Laws are only useful if violations are detected. The Sovereignty Protocol includes self-assessment loops — periodic evaluations where agents audit their own recent outputs against the laws they are supposed to follow.
A self-assessment loop works like this:
- The assessor agent reviews a sample of recent work outputs
- It checks each output against the active law set
- It identifies any outputs that deviate — in tone, in scope, in format, or in content
- It files a structured assessment report with specific examples and a drift score
- Operators are notified if drift exceeds a configured threshold
This is not the agent judging itself subjectively. It is a structured, law-referenced audit with explicit criteria. The assessor uses a different model instance than the agent being assessed, so it has no stake in the outcome.
Agent Roles, Flows, Templates, and Skills
Laws define what agents cannot do. Roles, Flows, Templates, and Skills define what they can do:
- Roles — scoped identities with specific responsibilities (e.g. Researcher, Writer, Auditor)
- Flows — repeatable multi-step procedures for common tasks
- Templates — structured output formats that ensure consistent, parseable results
- Skills — discrete capabilities that extend what an agent can do (browser control, file operations, API calls)
Together with laws, these form the complete governance surface of the Agent Brain — the part of the platform that defines the character of your AI workforce, independent of which model you are running under the hood.
The Free Tier Unlock
Agent Laws and the governance core are available on the Free tier. This is intentional.
We believe that the governance layer should not be paywalled. If you are using AI agents in any context — personal, professional, or experimental — you deserve constitutional rules that actually work. The service features (crawling, autonomous workflows, the model hub) are what we charge for. The architecture that keeps your agents honest is something you should be able to rely on from day one.
Start with the Agent Brain. Define your laws. Build your roles. Then add the service layer as your use case demands it.
The foundation is always there.