Standards
The agent lane works in conjunction with the standards. It is not a standalone hack. Each layer of the stack solves a problem the ad itself should never try to solve alone.
The stack the lab builds on
| Layer | What it provides |
|---|---|
| FAISC (Foreign Artificial Intelligence Site Connectors, Project Moosed's published standard) | An open layer any agent can read (manifest and feeds), and a connected layer where the agent is attached to the person's own account with a human-confirmed grant: defined scopes, an audit log, and revocation. Permission, audit, and revocation for agent access come from here. |
| Web Bot Auth and signed agents | Agent identity and authentication: a serving decision can branch on a verified agent identity instead of a self-declared user agent string, and every take can be logged against that identity. |
| ACP / AP2 / UCP (agentic commerce protocols) | Discovery, checkout, and signed payment mandates for agent-mediated commerce, as they mature. The lab's model sits on top of them: they standardise the transaction, the lab works on the advertising unit and its disclosure and billing rules. |
| Structured schemas and merchant feeds | Offer truth: prices, availability, inventory, and terms stated in machine-readable form that an agent can verify against the human-facing page before it presents anything to its person. |
| Crawl and access policy (robots and content signals) | Which agent classes may read what, at what rate, so a publisher can welcome its customers' agents without opening the door to unbounded scraping. |
White Hat rules for agent advertising
These are design requirements in the lab, not aspirations. An agent ad system that breaks them is, in the lab's frame, the Black Hat version of the same idea.
- Disclosure is required. When an agent presents an offer to its person, it must say that the offer arrived via a sponsored feed, and name the sponsor. The person trusts the agent, not the advertiser, and the agent lends that trust when it speaks. Hiding sponsorship spends trust the agent did not earn.
- Data-only ad schemas. An Agent Ad is typed data fields, never instructions to the agent. Content in an ad feed is untrusted data for the receiving agent, never a command channel. An ad that carries hidden instructions aimed at the agent is a prompt-injection attack, not an ad.
- No deceptive structured data. A price, availability, or term in a feed or schema must match the real offer an agent can verify. A feed that lies is filtered, and should be.
- No undisclosed pay-to-rank. If an agent or its platform is paid for placement inside the agent's own recommendation, that is a material connection and must be disclosed per offer. The agent may be a channel, or it may be a fiduciary recommender; it must say which.
- Human approval thresholds. The agent presents; it does not silently spend. Above a person-set threshold, explicit human approval is required before redemption.
- Bound, single-use, expiring instruments. Offer tokens are minted for one person, one agent, and one product or task, with a short expiry, server-side validation at redemption, and burn on use. A code that cannot leak cannot be traded.
This site's own agent layer
The lab eats its own cooking. The open layer of this site, live from version 1, is:
| Endpoint | What it is |
|---|---|
/manifest.json and /.well-known/agent-manifest.json | Site manifest: name, purpose, public feeds, and publisher (Project Moosed). |
/feed.json | Publications and updates feed (JSON Feed format): charter and paper entries, dated. |
/llms.txt | Plain text summary of the lab, its model terms, and its pages, for language models and agents. |
/sitemap.xml | The five public pages of the lab. |
| JSON-LD on the home page | Organization and WebSite structured data naming the lab and its publisher. |
This site carries no advertising, no ad serving, and no analytics trackers. The agent layer above is a publications layer, not an ad feed.