Repository navigation
Open
Description
Overview and problem statement
The registry is meant to be where a buyer, a seller or an agent finds and checks AdCP agents. Today it does not play that role:
- The public page builds its list in the browser, so crawlers and AI agents see no entries.
- Only members can list an agent, and new entries default to members-only. The public catalog shows 36 agents.
- Production, sandbox and demo endpoints sit side by side with nothing to tell them apart.
- Compliance results exist but are hard to read: 10 passing, 19 degraded, 7 failing, and 14 results older than 30 days with no stale marker.
- Nothing in a trade requires a registry lookup, so nobody has a reason to keep a listing current.
Third-party maps of agent interfaces now document who claims to support AdCP. Only AAO can say who was tested, and only AAO can answer "who is this agent and who authorized it". The registry should lead with those two answers.
Proposal
Three tracks. Child issues are listed in the order we suggest reviewing them.
| Track | Issue | Needs |
|---|---|---|
| Open | #9058 Open registry listing to non-members | Executive Committee decision |
| Open | #9061 Registry page readable without JavaScript | Engineering only |
| Trusted | #9059 Environment tag on registry agents | Engineering only |
| Trusted | #9060 Evidence badges on every registry listing | Working group review |
| Trusted | #9062 Listing quality checks | Engineering only |
| Trusted | #9063 Ecosystem roles and reach | Working group review |
| Used | #9064 Agent identity lookup for sellers and buyers | Working group review |
| Used | #9065 Registry where agents and builders already are | Working group review |
| Used | #9066 Search agents by what they offer | Working group review |
The public conformance feed and registry metrics are covered in #9060 and #9065.
Requirements and rules
- Open. Listing does not require membership. The owner of an agent decides its visibility.
- Evidence only. Badges, filters and ordering come from evidence that anyone can earn the same way. Membership tier, sponsorship and hosting provider play no part.
- Dated and sourced. Every fact on a listing carries its source and the date it was last checked. Unknown is shown as unknown.
- Neutral. Gateways, marketplaces and independent agents are treated by the same rules. Role definitions are set by the working group.
- Machine-first. Everything on the page is available through the API and readable without JavaScript.
Acceptance criteria
- Each child issue is accepted, amended or closed by the working group.
- The Executive Committee decision on Feature: Open registry listing to non-members (catalog, API and registry page) #9058 is recorded.
- Registry lookups per day and the count of conformant production agents are published.
Open questions
- Which working group owns the registry roadmap?
- What is the target date for Feature: Environment tag on registry agents (production, sandbox, test) #9059 and Feature: Registry page readable without JavaScript (server-rendered list and agent pages) #9061, which need no policy decision?
Out of scope
- Membership pricing.
- Changes to the compliance storyboards.
Metadata
Metadata
Assignees
Labels
No labels
Activity
EvgenyAndroid commentedon Nov 17, 2026
Disclosure first: I run the Agentic Ecosystem Atlas, which I think is one of the third-party maps mentioned above. I'm also in AAO's Signals and Measurement working group.
I support the split. The Atlas records what vendors document. It can't say what was tested, or who authorized an agent. I'd rather cite AAO for both than build a weaker version of either.
Two notes from the mapping work:
#9060: the public conformance feed is the piece we'd use. For a third party to cite a result, each one needs
track,adcp_version,environment, the suite or SDK version it ran against,checked_at, the issuer, and a stable agent identifier. Please keep past results rather than overwriting them, so a stale result can still be cited with its date. We'd import these as protocol conformance only, labelled with issuer and date, and never as evidence that an agent is safe to spend through.#9063: buy-side reach is rarely published.
adagents.jsonproves reach on the sell side only. Until there's a common proof for buyer agents, I'd keep buy-side reachself_declaredwith a source link, and show "not published" as its own state, separate from zero. On open question 3: some operators need more than one role. A primary role plus secondary roles has worked for us.On #9059: my sales agent is a ready test case.
sell.nofluffadvisory.comsells media on nofluffadvisory.com and would be taggedproduction.sandbox.sell.nofluffadvisory.comis its sandbox, which is the pairproduction_counterpartis meant for. Happy to tag both early.lukasz-pubx commentedon Nov 18, 2026
From the buy side, one gap I'd like on this list: onboarding. The epic leads with two answers, who was tested and who is this agent and who authorized it. A party about to trade with an agent it has never met needs a third: will we work together, and what is needed first. It also bears on the last point in the problem statement, that nothing in a trade requires a registry lookup.
AdCP 3.3 already publishes the inputs (experimental, in 3.3.0-rc.0; #8113 and #8114, shipped in #8170). Sellers declare an onboarding posture in
get_adcp_capabilities.account.agent_onboarding: a mode (open,application,invite_only) per billing party and for sandbox, with requirements and a starting URL. Buying agents declare theirs in their operator's trust.json. Nothing reads them across the ecosystem yet. The closing note on #8118 says "The registry crawl (R1 follow-on) was never milestoned and is not tracked yet." I've filed it as a related issue, #9089, in the same format as the issues listed in the epic, and suggest it for the Used track.What the registry can compute differs by direction, and the issue says so:
/verificationand/storyboard-status), so "which buyers' published requirements does this seller meet" becomes one query across the catalog.grantfor a named brand, using the trust.json lookup that already backs trustjson.org. It cannot see an existing billing relationship, whichopenbilled tiers require, or the human steps (billing_entity,contract,credit_application). Those show as not checkable, never as met.The match also gives sellers a reason to publish a posture at all: a seller whose sandbox is
openshows up in buyers' filters.How it touches the related issues in the epic:
conformancerecord lands, please keep the badge claim names the registry already returns (role,verification_modes,verified_specialisms,adcp_version). A buyer'saao_verifiedrequirement uses exactly those names.brand.jsonauthorized_operatorscounts only where the brand's parent organization has never published a trust.json. Grants name operator domains, and name agents only when they narrow withagents[].self_declared, and the absence of a grant says nothing.Later: competences. Onboarding answers whether two agents can trade. The next question is what an agent can reason about inside a task: whether a seller counter-proposes on
refine_proposalsor only accepts and declines, and whether it can work with a free-text brief. An RFC of mine, #6287 (decision tracked on #6508), proposes that agents declare this as a name, a task scope and a description. It is not in the spec, so #9089 does not depend on it, but the registry is where it would pay off:refine_proposals, including sellers it has not contacted yet, instead of asking each seller it already knows. That works only if competences are readable without credentials, as capabilities are.evidenceslot on each competence and leaves its shape open until there is a record worth pointing at. Theconformancerecords in Feature: Evidence badges on every registry listing (registry page, API and public feed) #9060 and the signed evidence records in Production evidence program: machine-observed AdCP implementation matrix #7068 are that kind of record. A competence with nothing behind it stays a declaration, which is the line the RFC: pre-proposal competence discovery for conversational agent routing #6508 triage draws: "a descriptive competence must not be mistaken for a supported task, certification, or authorization claim".metin the match says whether it rests on a declaration or on evidence. That already matters forsandboxandexperimental_feature, and it lets a competence requirement join later without changing how the match reads.The match follows the epic's rules: public documents only, dated and sourced, and advisory. It never approves or blocks a trade.