Method · dataset v2026-11-15.30
How the Atlas decides
The Atlas records what public documentation says an agent can do through one exact interface. It is documentary scope: not account entitlement, not tested behaviour, and not a recommendation.
1
Four questions, four states
Every interface gets the same four questions. Read: can it return data? Write: can it create or change something? Optimize: can it steer the platform's own bidding, budget or optimization settings? Activate: can it put campaigns or audiences live?
- Documented: the vendor's own material describes it.
- Limited: documented with a stated narrowing (some objects, a draft that a person publishes, an approval step).
- Excluded: the documentation rules it out explicitly ("read-only", "cannot modify"), or an exhaustive contract (a complete tool or path list) has nothing that does it.
- Unknown: the evidence we read does not settle it. Unknown is not no.
The negative-evidence rule. A tool missing from one guide is unknown, not excluded. Excluded needs a rule-out or a complete contract. This is why Sovrn's Curation MCP and WorkMagic's MCP show Optimize as unknown since 14 Nov 2026: their guides neither document nor rule it out.
2
Task answers are checked on the exact interface
A task is stricter than a capability. An interface qualifies only when the same interface documents the capability and its write or activation reaches the task's object. A different API at the same company does not count.
- Create or change a paid campaign or media buy: needs Write documented or limited, and a reviewed write object of the right kind.
- Change bids, budgets or pacing: needs Write documented or limited, and a reviewed write object of the right kind.
- Put an audience live: needs Activate documented or limited, and an activation effect that delivers an audience.
What does not count. A forecast, scenario, plan, draft or report that mentions a budget is analysis, not an applied change (WorkMagic's budget optimizer forecasts; it does not count as changing bids or budgets). Audience membership is not campaign attachment (Kayzen's audience API builds lists; it does not put them on a campaign). A protocol route run by someone else is not native support (Google Ad Manager reaches AdCP via Prebid's Sales Agent). Documentation-only servers, in-app agents and announcements are reference only.
Where scope comes from, strongest first: Reviewed operation (objects and approval mapped by hand); Typed objects or effects from the record; Reviewed use-case tag only (objects not typed); Scope not mapped. Interfaces whose objects are not mapped are never counted as a yes or a no.
3
Access badges are derived, and say so
The badge comes from the reviewed access note by transparent rules, read clause by clause; a negated clause ("no allowlist", "self-service access not verified") never counts as a gate. Tiers: Open: no account needed; Any customer account; Approved partners or by request; Customer account plus an entitlement or enablement step; Invite, waitlist, beta or staged rollout; Not established. The note itself, on the record, outranks the badge. Who may get access, what the account needs, how you sign in and who approves each action are separate facts.
4
Freshness: three clocks, kept apart
- Claims reviewed
- The date a person last re-checked the record against primary sources: 431 of 477 interfaces so far. Only a dated check moves it; a new release, a watch run or a probe never does. Records not re-checked yet show their research date instead. Review due after 90 days.
- Sources watched
- A weekly script reads the cited pages of agent interfaces and of anything that can write or activate, and fingerprints the sentences about capabilities, lifecycle and approval. Last run 13 Nov 2026: 373 pages. When a cited page changes after the last review, the record shows "Source changed, review pending" until a check dated on or after the change. The states stay as they were in the meantime: a changed page is a reason to look, not a verdict. A cited page that returns HTTP 404 or 410 shows "Cited source unavailable". Pages that render in JavaScript give a thin fingerprint, so a change there can be missed.
- Endpoint probed
- A public, credential-free MCP handshake (initialize and tools/list, no tool calls) against the endpoints the Atlas records, at most weekly; last run 14 Nov 2026. Outcomes: Public tool list answered; Endpoint answered the MCP handshake; Endpoint live; sign-in required; Recorded address did not answer as MCP; Recorded address unreachable. A failed handshake is an observation about an address on a date: never an unsupported capability, a revoked account or a confirmed outage.
5
Setup guides
For 6 high-demand interfaces the company page carries a setup guide: who provides it, where it is hosted, the published endpoint, sign-in, the official example config, a first read-only call, and whether a sandbox and pricing are published. Every line comes from the vendor's own pages on the date shown; config text is copied with its placeholders and holds no credentials. A first read is an account-data call under your own sign-in. A guide never widens access, approval or the capability states, and an interface without one simply says its setup is not mapped yet. The same facts are in the open index (setup, schema at atlas-index.schema.json) and in the MCP server's get_company.
6
Evidence depth and difficult cases
For 44 high-demand interfaces every claim (read, write, optimize, activate, access, sign-in and date) is tied to the one vendor source that decides it, with a short paraphrase and the date it was read; a claim the public pages do not settle says so. Their task answers come from the published tools and objects, and name the operation where the vendor publishes it. The sources and paraphrases are on the interactive record; the open index and the MCP server say which claims are covered (evidence_depth).
Dates keep their precision (a month-only listing stays month-only) and say what they are: first release, general availability, beta start, announcement, directory listing, or the earliest dated source we found. Sign-in says when the method is only implied (auth_basis: named, inferred or unknown). The weekly watch covers agent interfaces and documented writers; the Sources and changes tab shows how many of those it covers, not "all".
Difficult cases. qa-cases.json lists 28 hard calls (one interface not inheriting another's authority, a modelled budget that is not an applied change, excluded versus unknown, derived access and sign-in, date precision) with the expected answer, the reason and the Atlas's current answer from the live engine. The build fails if any disagrees. It is a bounded set for checking the reasoning, not an accuracy rate.
Agentic Maturity Index. The index scores every company 0–100 from these same documented facts (agent reach 20, capability depth 40, task authority 25, access 10, evidence 5). Unknown counts zero. Vendor momentum counts only interfaces with a dated launch, and weekly snapshots are scored with the same formula, so a revision always means the data changed.
Comparisons. Five side-by-side pages are built only from interfaces reviewed claim by claim; their answers come from the same engine, and a test fails if a page's framing stops matching the data it describes.
7
Corrections
Anyone can suggest a correction; the company pages prefill the record. A primary source is what changes a state: the vendor's own documentation, release note or registry entry. A vendor's email can point us to a source; it does not change a state on its own. Accepted corrections are dated, listed with what the record said before and after, and kept. When one of our own corrections turns out wrong it is marked withdrawn, not deleted. Every change has a stable ID in changes.json and the RSS feed.
Automated versus reviewed. Scripts fingerprint pages, probe endpoints and derive badges. People read the sources and set states, task scope and corrections. No script changes a state.
Cite as: No Fluff Advisory, Agentic Ecosystem Atlas, method, dataset v2026-11-15.30, https://nofluffadvisory.com/agentic-atlas/method/