orange tomato
Page 1/10

WHO AUTHORIZED THE AGENT?

A companion notebook to the deep dive — payments, purchases, permissions, installs
YOU authorize
→
the AGENT acts
→
a COUNTERPARTY decides
The question on every surface: what did I actually authorize — and who answers when it acts?
  • Four surfaces, seven absorbed stories, one pattern.
  • Every boundary is real. Every document names a human back.
  • What the record prevents: less than the phrase "permissions and audit trails" suggests.
10 pages · compiled 2026-11-08 · sources and dates on page 10
orange tomato
Page 2/10

1. THE AUTHORITY CHAIN

You (account holder)
→
Agent (runs under your name)
→
Counterparty (its own terms)
A delegated action crosses TWO gates in sequence:
  • Gate 1 — you authorize: a limit, a consent prompt, a permission, an install click.
  • Gate 2 — the counterparty accepts: merchants, platforms and registries each decide separately whether to serve the agent at all.
  • Your instruction is only half of the transaction. If gate 2 stays closed, there is no relationship for the agent to step into.
consent gate
merchant gate
permission prompt
tool call
Enforcement lives at a named moment: at authorization, at checkout, at the prompt, at the tool call. Not "somewhere in the product".
orange tomato
Page 3/10

2. PAYMENTS — "AGENTS PREPARE, PEOPLE CONFIRM"

  • Two layers. The connection layer prepares: the agent drafts, the person confirms, money movement isn't exposed. Agentic Payments — a separate opt-in product — lets an agent actually pay.
  • Still a pilot (checked Nov 8, 2026): "Agentic payments is rolling out with pilot partners and customers."
  • The boundary is set before the agent acts: budget, period, merchant categories, card type, and one of three oversight modes.
  • Enforced at the moment of spending: "Card limits are enforced at authorization."
  • Named accountability: every agent paired with a user; every payment names both; no self-raised budgets; one action revokes.
person sets boundary
→
agent prepares
→
limit checked at authorization
→
human confirms
The audit trail is a photograph, not a lock. It records; it does not stop a bad configuration, a mistaken approval, or a dispute. Merchants may decline unknown automated buyers; read data flows to the model provider; the controls are vendor-documented.
orange tomato
Page 4/10

3. PURCHASES — THE SECOND GATE HAS A RULEBOOK

  • Walmart Agentic Access Terms (Oct 21, 2026 version — unchanged on Nov 8): agents must identify and "must not disguise the fact that it is an agent"; explicit named-agent consent; a Human Consent Gate before every checkout outside the Agent Gateway.
  • Liability is named: the retailer is not liable for decisions made by an agent; user and deployer are responsible; aggregate liability capped at $50; human statutory rights unchanged.
  • Amazon: identified agents tolerated (`Agent/[name]`), with a reserved right to limit, block, or order an agent to stop.
  • Terms-of-use layer (attributed — trackers, not verified at the page): authorized-agent actions within your permissions are treated as your own actions.
$50
is the retailer's exposure. Yours is unlimited. The returns flow still runs — and the account the order sits on is yours.
Gateway mechanics (tokens, signed mandates, guardrails) are terms-described and in early beta — not observed in use.
orange tomato
Page 5/10

4. PLATFORM PERMISSIONS — WHAT "ALLOW ALWAYS" ALLOWED

  • Muse (Meta's personal agent) reached ~10 million users by Oct 21. The permission prompt is the whole authority layer: allow once / this task / this site / always / deny.
  • Shape 1 — scope by time and place, not by consequence. "Allow Always" on messages covered future acts nobody named: sharing an address, accepting an offer, replying as the seller.
  • Shape 2 — the agent that couldn't explain itself. A disputed message-access account; what both sides agree on: the agent gave an incorrect explanation of its own internals. An audit trail is only useful if the agent can explain it.
  • Shape 3 — unguessable is not private. Share links built from conversation titles sat in public DNS; 6,800+ hostnames findable by name (single-source count, attributed). Sharing paused Oct 20, restored Oct 22 — the fix is the vendor's claim.
Habit the help pages suggest: check the activity log after anything consequential — that is where an unexpected action surfaces, usually after the fact.
orange tomato
Page 6/10

5. EXTENSION — THE AUTHORITY YOU INSTALL

  • A skill is a package: instructions, trigger description, scripts, dependencies, permission grants — five surfaces a download count cannot show.
  • On Claude Code, a project skill's `allowed-tools` can pre-approve tool use for its turn, even in a folder you've never trusted (runtime docs warn in as many words).
  • Scanners help and self-bound: static only, capped ingest, coverage that can be incomplete. "Clean" is not "safe".
FigureCountsSource
42,447 → 31,132skills collected → analyzed (two marketplaces)Liu et al., 2026
26.1%analyzed skills with ≥1 vulnerabilityLiu et al., 2026
5.2%high-severity patterns suggesting malicious intentLiu et al., 2026
2.12×likelihood for script-bundling skillsLiu et al., 2026
The habit: read the whole package · test the trigger both ways · review the permission grants first · pin the version · run a real task with and without it.
orange tomato
Page 7/10

6. THE PATTERN — BOUNDARIES, PHOTOGRAPHS, PEOPLE

  • Boundaries: real, narrow, enforced at a named moment — on every surface.
  • Photographs: the record of what happened is everywhere; the question of whether it should have is answered nowhere. And sometimes the agent misreads its own photograph.
  • People: responsibility documents converge — network rules hold the cardholder to the purchase "as if" they made it; retailer terms name user and deployer; the payment platform pairs every action to a person; the skill ecosystem asks you to read the package.
permission policy
what you told it
technical boundary
what is impossible
visibility
what the log shows
Three different people maintain these three: you, the vendor, the database.
"An agent you authorized can still be a stranger to the store." The systems are not failing to define responsibility — they are defining it, consistently, downward.
orange tomato
Page 8/10

7. WHAT YOU CAN INSPECT

  • Which layer am I on? Preparing ≠ executing. Connection ≠ payment. Listing ≠ checkout.
  • Where is the boundary enforced? At the moment, or only in the terms?
  • Who is named? Paired user, account holder, deployer. If the answer is nobody, read it again.
  • What does the record cover? What it shows — and what it cannot prevent. Can the agent explain its own access accurately?
  • What does the counterparty reserve? Block, throttle, change terms, decline automated buyers.
  • What is the revocation path? Prefer one action. Test it once before the stakes are real.
  • Who answers when it breaks? Dispute route, return process, cap. Statutory rights usually survive the agent layer — the document should say so.
If any answer is "the demo didn't say", don't treat the action as routine.
orange tomato
Page 9/10

8. WHAT CHANGES THE JUDGMENT

  • Personal Agent Protocol ("Poppy") — Meta + Sierra with Shopify, Stripe, Walmart and others. v0.1 published Oct 9, v0.2 Oct 24, 2026 (drafts before 1.0). A company publishes a discovery file; the agent opens a session with the scopes you granted (e.g. `poppy:read`); either side can bring in a human. Payments are not covered yet. Vendors sell the machinery — framing disclosed.
  • Smaller clock hands: Walmart's Gateway beta → general availability; Agent Profiles promised; Pliant's pilot → GA with independent testing; Muse's product state moved twice in three days in October.
  • Not here yet: an operative dispute/liability rule. The forecasts (a network rule assigning agent-purchase liability; a state law accepting agent claims) are predictions, to revisit.
If the standard ships and the pilots turn general, the checklist shrinks — identity and scope stop being things you personally verify. If they stall, the checklist is the product, and its first line does most of the work.
orange tomato
Page 10/10

9. SOURCES & METHOD

  • Method: desk research and source audit only — no purchases, payments, accounts, installs, or scanner runs. Reading pages is not running anything.
  • Dated readings (Nov 8, 2026): Walmart Agentic Access Terms — still the Oct 21, 2026 version; Amazon Conditions of Use (Agent Terms present; page undated); Pliant product and docs (pilot unchanged); PAP announcement and spec v0.2; Liu et al. (arXiv:2601.10338) retrieved.
  • Carried material from seven earlier pieces, each re-cited to its primary source; none of the retired articles is cited in the deep dive.
  • Attributed, not verified: Visa rulebook content (via a dated analysis); the terms-of-use "your actions" layer (via trackers). Claroc's findings attributed to a vendor with commercial interest. WorkBuddy metrics disputed. Evidence reviews of the absorbed pieces were not independent.
  • Open at compilation: Pliant dispute/reversal mechanics; the iMessage dispute (both sides); the Oct 22 share fix is the vendor's claim.
Companion to "Who Authorized the Agent?" — the article carries the full source list with links at the point of each claim.