Polis Protocol (Multi-Agent Teams)

Set up a self-optimizing, multi-vendor AI agent team in a folder of Markdown. Each agent is a citizen with a signed capability card; tasks are contracts routed to the best performer by a learning bandit; settled contracts produce lessons that compound into team memory.

Polis Protocol treats a project as a small city of agents from any vendor (Anthropic, OpenAI, Google). It defines a Register (identity and capability cards), Contracts (structured tasks with learned routing), a Chronicle (lessons that feed back into routing), and Amendments (agents can change the protocol itself). Work is routed to whichever model is actually best, and the team measurably improves over time — all in plain Markdown files.

When to use

Use when several AI agents collaborate on one project and 'who should do this' matters, when you want work routed to the best model and a team that improves with use, or to register a citizen, open/claim/settle a contract, file a lesson, or propose an amendment.

Examples

Bootstrap a polis

Create the protocol folder and register citizens

Set up a new Polis Protocol folder for this project and register me and two other agents as citizens with capability cards

Route a task contract

Open a contract and let routing pick the best agent

Open a task contract for refactoring the auth module and route it to whichever citizen has the strongest track record