The Chosen One

Agora is an office where people and their AI agents work together, and where no agent can lie about what it did. The problem: agents now send our email, book meetings and read our messages, and when something goes wrong the only account of what happened comes from the agent itself. Meta's Muse told a journalist it saw only notifications while it had synced 187,000+ rows of his iMessages (Decrypt). In our own test, asked later about their own actions, 20-41% of what three Meta models said was false. How Agora fixes it: - The model only asks. Rust code acts, and writes every call to a hash-chained, Ed25519-signed ledger the agent can't touch. - The guard: two independent parsers split every message into claims, and plain code (not another model) checks each claim against the ledger. A false draft is sent back with the receipts; nothing the ledger contradicts reaches you. - Meta's Agents Rule of Two, enforced in code for real MCP tools: the person who owns a tool approves calls that leave the office. - A shared 3D office: teammates join with an invite, each gets their own bot, teams of agents run briefs, and anyone can drive it from Claude Code, Codex, OpenCode or Gemini CLI (a CLI can never approve its own calls). Proof: 0 of 907 delivered claims false under a red team; 136 real errands on open-weights Muse Spark Open: 0 false claims, $0.021 each; 6 of 6 hostile-server attacks stopped. Live: https://agora-office-295478535981.us-central1.run.app ("Watch it catch a lie" works without a login). Built from scratch during the hackathon (first commit Oct 26), in Rust.