The AI era does not need one flow for people and another for Agents. It needs shared rules: sellers publish facts, Agents prepare, humans approve, settlement stays visible, and every state leaves evidence. WebAZ is testing that stack:
- An AI agent writes 80% of a patch. A maintainer changes 20%, runs the tests and ships it. Who contributed? A useful record separates requester, executor, reviewer and accountable signer. WebAZ's boundary: AI may execute; a Passkey-bound human remains accountable.
- Contributor onboarding should not start with “create an account.” It should start with: here are the open problems, boundaries and evidence expected. WebAZ lets anyone inspect public tasks or suggest one without an API key. A suggestion is not yet a contribution.
- “Payment succeeded” is not “transaction complete.” Did the merchant acknowledge it? Was inventory committed? Did fulfillment start? Is the outcome unknown or safe to retry? Agent commerce needs explicit state ownership, not one green checkmark.
- An MCP server is not interoperable just because it connects. I tested WebAZ's public shopping endpoint with FastMCP 3.4.7: schema, is_error, concise content, and structured_content all passed in one read-only call. No account, order, or payment access.
