What is an AI agent actually authorized to do? An agent having access to a tool doesn't necessarily mean it was authorized to perform this specific action. So I built Agent Authority around five primitives: Passport — identity Contract — authority Gate — enforcement Box — execution boundary Ledger — evidence The developer preview includes signed identity, task-scoped contracts, bounded delegation, approvals, replay protection, deterministic authority scoring, and portable evidence. I’m not launching this as a finished security product. I’m launching it so developers can use it, challenge it, and find what I got wrong. If you build agents, MCP servers, automation, or agent-to-agent systems, I’d love to know: What authority does your agent have today that you wish you could make explicit?

Agent Authority

Give every AI agent a permission slip

Give every AI agent a permission slip