Skip to content
Citizen Infrastructure

Constitutional agents

There’s a deep question at the heart of any digital governance system: should the rules be code or interpretation? Should a deterministic program decide who gets banned and which proposal passes — or should something read the community’s rules and apply judgement? Modular politics assumed the former. Nathan Schneider’s more recent work explores the latter, and calls it constitutional agents.

A plain-language constitution, interpreted by an agent

Section titled “A plain-language constitution, interpreted by an agent”

The idea: instead of freezing governance into code, a community writes its rules as a human-readable constitution — in plain language anyone can read and edit — and puts an AI agent at the centre of the system to interpret and apply it. Schneider has been testing this with a Mastodon bot: it holds a plain-text constitution, and when a member flags a post, the constitution might say “this is removed if five other members agree,” or “this decision is delegated to a chosen group.” The agent runs the process, keeps a ledger, and acts.

The appeal is accessibility. With code, implementing a new governance practice — maybe one from your own culture — means finding a programmer. With an interpreted constitution, you could just describe what should happen in words. That hands less power to technical experts and more to the community itself.

The catch

Section titled “The catch”

Schneider is explicit about the risks: today’s LLMs are “profit-driven, fundamentally extractive tools,” and pure code-based systems (think DAOs) have shown how rule-by-code gets gamed by whoever’s best at gaming it. He wouldn’t hand an agent ultimate power. But as a way to make self-governance legible to ordinary members — and to detect the spirit, not just the letter, of the rules — it’s a striking direction. It also rhymes with a broader shift: even AI labs (he points to Anthropic’s “constitution” for its Claude models) increasingly govern models in human terms rather than pure code. See also civic AI.

Sources

Section titled “Sources”