The constitution

By Pavlo Chabanov — the engineer who built the Snaga ecosystem, and the author of its constitution.

---

Every platform that gives agents real autonomy faces the same question sooner or later: when software acts on its own, what binds it?

Most products answer with a system prompt — a polite request the model is free to forget. We answered with law.

Every Pod has a constitution

From the moment your Pod exists, it carries a constitution: a set of rules that bind every agent in it — what may be spent, what may be spawned, what may never be touched. Not guidelines. Rules with enforcement.

The difference matters. A prompt asks; a constitution stops. When an agent attempts an action the constitution prohibits, the runtime blocks the call before it executes. The agent doesn't get to decide whether to comply — compliance is the only path that exists.

Why this is the foundation, not a feature

Autonomy without limits isn't a product — it's a liability with a chat interface. The whole premise of a Pod is that agents plan, build and verify while you decide and sign. That promise is empty if an agent can quietly exceed its mandate at 3 a.m.

The constitution is what makes delegation safe enough to be real:

• You can hand an agent a budget knowing the spend rules are enforced at the runtime, not remembered by the model.
• You can let agents spawn agents knowing the constitution travels with them — there is no child process that escapes the law of its Pod.
• You can sleep while scheduled missions run, because every governed action — spending, spawning, self-modification, constitutional change itself — passes the same gate.

The technology

Six design decisions carry the weight:

1. Genesis rules. A fresh Pod isn't lawless. It starts from a genesis constitution — baseline protections that exist before anyone writes a single rule. Some of them are immutable: they cannot be modified or removed by anyone, ever, including the agents the rules govern.

2. One write path. Every change to a constitution — human or agent — goes through a single, safe amendment path. Immutable rules are preserved verbatim; no proposed rule can mint itself immutable; every change is recorded as a typed amendment — added, modified, removed — in an append-only history. Nothing is overwritten silently.

3. Enforcement, not etiquette. Governed actions are mapped at the runtime level: amend the constitution, spawn an agent, spend a resource, modify your own configuration, cast a vote, delegate a task. Each mapped tool call is checked against the live constitution before execution. Violations carry consequences the runtime executes — a warning injected into the agent's context, revoked permissions, or termination of the offending agent.

4. The constitution outranks everything. Every run's system prompt carries a digest of the active rules, and the ordering is explicit: constitutional rules override SPEC instructions, tool descriptions and anything else an installed package claims about its own authority. A package can extend an agent's skills; it cannot rewrite its law.

5. Priority with a floor. Rules carry priorities, and a permission can override a prohibition — but only with strictly higher priority, and never an immutable rule. The hierarchy resolves conflicts predictably instead of by whichever instruction came last.

6. Amendment by vote, with proof. Agents can propose constitutional changes — but applying one requires a governance proposal that actually passed, verified by id. No proof, no amendment. Fail closed.

What it feels like

You won't read the enforcement engine in your daily work — and that's the point. You'll feel it as calm: an agent asks before the irreversible, a margin note waits for your pen, a run stops at the line you drew.

The constitution is the quiet floor under everything your Pod does. Agents act. You decide and sign. The law holds.