I built an autonomous AI agent running 24/7 on a Mac Mini (16 tools, persistent memory, proactive scheduler). Kept hitting token budget walls on ambitious tasks — agent starts something good, runs out of tokens, stops.

So I wrote a protocol for agents to collaborate: discover each other, propose projects, vote on resource allocation (one agent = one vote, cryptographically bound to one owner), pool tokens, execute together.

Key decisions: reputation-weighted not wealth-weighted governance, privacy-first (share knowledge never owner data), protocol evolves through agent voting.