This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of both Anthropic and OpenAI and also tests Google models. Poppy, in item 1, sits on top of MCP, a specification Anthropic originally created, and we disclose that. We run commerce, apps and an AI influencer ourselves, so in all three stories we stand on the side that has to write the document — which is why we add no judgment to any company's approach and only relay what is written in the documents and what is not. Figures announced by companies are marked "announced by the company," and where something comes from a single source we name that source. The health-related statements in item 3 are not medical advice. This article is not an investment opinion on any company.
In NO.43 we wrote that "customers cannot bring their own agent," in NO.45 that United announced a "ban on bookings by personal agents," and in NO.49 that AT&T set a threshold of "no autonomy before accuracy in the mid-to-high 90s." This week's three stories are the next step — writing down "what is allowed" in a document before the action happens. The store writes one file before the agent arrives, the company writes one registry line before running the agent, and Apple writes three conditions before the answer goes out. We separate fact, interpretation and "if it were us."
1. Poppy 1.0 — the store writes "what an agent may do in our house" in a single file (fact, company announcement)
Sierra (the AI customer-service company led by former Salesforce co-CEO Bret Taylor) released "Personal Agent Protocol" — nicknamed Poppy — version 1.0 on its official blog on November 13 (co-written by Bret Taylor and Clay Bavor). It first announced the protocol with Meta on October 6, published a draft on October 9, and wrote that it "tested the drafts against real websites, APIs and agents for five weeks." We read the specification itself (personalagentprotocol.org). This is how it works — a company places one JSON document at `/.well-known/poppy.json` on its own domain. Inside are `organization` (name, domain, logo), the doors an agent may enter (`agent.protocols` conversation endpoints, `apis` — OpenAPI or MCP, `web` — a browser-session endpoint, `tasks`, `extensions`), and policies (`auth.disabled` to turn off sign-in, `web.disabled` "do not browse our website," `web.unscoped` whether browser sign-in means full account access). There is a `scope_hints` field that marks each scope as `irreversible` (an action that cannot be undone) or `personal_data` (involving personal data). Sign-in uses OAuth 2.0 (PKCE, device sign-in RFC 8628) and sessions use a signed `Poppy-Session` token. The spec describes itself as "1.0 is 0.4 plus fixes" and hands areas not yet covered to a roadmap page. Sierra's post says Poppy sits on existing standards such as MCP, OpenAPI and UCP, that the commerce extension was co-authored by Shopify, and that there is an implementation kit with a validator.
Participation figures are the company's. To October's eight initial design partners (Meta, Genesys, Instinct, NiCE, Rocket, Shopify, Stripe, Walmart) sixty more were added for more than 100 (including Affirm, Delta Air Lines, Salesforce, Twilio, Wayfair and Zoom), and those actually live are Mercury (a bank — payment requests go through the account's approval policy), Synchrony and Turkish Airlines (loyalty members approve a business-class upgrade before payment). Two sentences are the center of this article. "Every Shopify store now automatically publishes a Poppy file, so merchants don't need to do any extra implementation" (Sierra) — Shopify powers stores in more than 175 countries; UCP handles cart and checkout, while Poppy handles "who are you and what may you do" before that. And "Meta's Muse automatically looks for Poppy first on every site it visits." TNW (November 13) put Muse's users at "more than 10 million" (one outlet), and added that despite the 1.0 number the spec is labelled a draft and may change in ways that break compatibility, that Visa and Mastercard are building separate agent-identity checks, and that Amazon wants shoppers to use its own agent instead of Muse.
Two lines of background. According to Practical Ecommerce (September 24), on September 8 — the day Muse launched — Shopify added Meta to "Agentic Storefronts," sharing merchant products with Meta by default (merchants can turn it off), while Amazon blocked the same Muse with a message that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use" (NO.43). The words of Sierra's co-founders relayed by CMSWire (October 7) are "consumers decide what access to give their personal agents, and companies set limits on what those agents can do," and Meta called it "a standard that, like email, no single party owns," but no governance model was published and the two companies' partner lists differed. The spec site's footer states Apache License 2.0 (personalagentprotocol.org, checked directly on November 18); no standards body or governing entity is named. Payment is not yet core but a "candidate extension."
2. Uber — "identity only answers 'who'"; each agent's permitted intents and risk tier are registered, then inspected at the moment of execution (fact, company announcement)
We read the Uber Engineering blog post "Securing AI Agents: Lessons From Real-World Deployment" (November 6, by Sean Tout, Max Zhurovich and Adi Raghavendra) directly, and InfoQ (November 15) summarized it. The starting sentence is this — "Traditional authorization can determine who is allowed to act, but it can't determine whether an agent's proposed action is appropriate in context." What it tries to stop is prompt injection, tool and memory poisoning, goal hijacking, behavioral drift and hallucination-driven actions (methods are not in the post, and we do not write them either). The structure has three layers. ① Before deployment: every agent is registered in the "AI Registry" with identity, owner, persona, permitted intents, constraints, dependencies and risk tier, and an automated analyzer called "Argus" scans code, prompts, configurations, dependencies, endpoints and tool-invocation paths for weaknesses and pre-generates persona and intent drafts for developers to review (MCP servers and tools go through the same process). ② At runtime: "AI Guard" is the policy decision point, running on small specialized models hosted locally, and "Agent Guard" hooks intercept actions at the boundaries of agent and tool execution and apply the verdict. What is checked is whether the action fits the persona, permitted intents and semantic constraints, whether it suits the user's request and role, and whether it makes sense given the whole session. ③ After deployment: an execution trace that links agent actions, tool calls and decisions in one line (sensitive data redacted, content included only when needed), and a device that "replays" recorded interactions against new policies before they are enforced.
The difficulties are written down too — the balance between false blocks and false allows (handled by shadow mode that only records without blocking, replay, and per-agent calibration), latency (parallel execution, small specialized models instead of large general ones), developer friction (Argus automation), and the fact that "the same rules for every agent" did not fit agents with different purposes and risks. The post closes with "security, accuracy, latency, cost, agent autonomy and developer experience can't be optimized independently" and promises follow-up posts (persona, context-aware guardrails, observability, intent filtering). There is not a single figure — no false-block rate, latency, cost or agent count. Nor does it say whether anything is open source. For reference, there is a report (Techstrong, August 10) that Uber released a separate tool called "ADR" in August — more than 50,000 agents, ten months in operation, "credential exposure was more common than prompt injection," all company figures — but neither that article nor the November post mentions the other, so we could not confirm the relationship between the two. We relay one reader's question from the InfoQ article as is — "the open question is how this works for agents you don't build."
3. Apple — attaching "clinical citations, detection of concerning questions, no training use" to health AI answers and asking 200,000 people to rate them (fact, company announcement)
On November 15 Apple announced the "Apple Health AI Usage Study" in its Newsroom and the same day put the "Apple Health AI Study App" (developer Apple Inc., free, iPhone only, iOS 26.0 or later) on the US App Store. We read both pages. The collaborator is OpenEvidence — a Miami company founded in 2022 that runs an LLM medical search engine for physicians; according to Wikipedia, as of December 2025 about 760,000 US physicians were registered and it was used for 18 million clinical consultations a month (company figures), and it holds content licenses with NEJM and JAMA (we do not cover its fundraising in this article). Eligibility is US residents aged 21 or over, an iPhone 15 Pro or later with Apple Intelligence enabled, iOS 26 or later, up to 200,000 participants. Participants ask everyday health questions, read complex medical information in plain language, and rate the quality of answers in surveys. Answers draw on medical sources such as NEJM, JAMA and Nature, and each answer carries links to peer-reviewed clinical citations. Those who wish can share resting heart rate, activity, sleep and lab results from the Health app to see how much personal data changes the relevance of answers. Three conditions are written in the post — "study data will not be used to train AI models," the app has guardrails "intended to detect certain concerning questions and suggest that a participant follow up with a provider or other resources," and "the app is not meant to diagnose or treat any condition, or replace the advice of medical professionals." Sumbul Desai, M.D., Apple's VP of Health and Fitness, said "health is deeply personal" and called the study "the most thoughtful way to understand how people engage with, comprehend, and evaluate health AI." The study's duration is not in the Newsroom post.
What the media added comes one item per outlet. TechCrunch (November 15) wrote that the study runs more than 12 months, that Health data is anonymized before going to "a secure third-party cloud provider," and that OpenEvidence deletes query and response data after 24 hours (these three are not in Apple's Newsroom post, so TechCrunch alone). The same article added that OpenAI said earlier this year it receives "more than 300 million health-related queries per week" (OpenAI's original is not reachable from our side, unconfirmed). TechCrunch gave the device requirement as "iPhone 15 or later," while Apple's Newsroom and Business Standard (November 16) say "iPhone 15 Pro or later," so we followed Apple's wording. iGeeksBlog (November 16) quoted Apple as writing that LLM-based consumer health tools "may at times offer responses that are incomplete or unclear," but we could not find that sentence in the Newsroom text we read, so we leave it as one outlet. The App Store privacy label read "User ID — linked to you (analytics, app functionality)," with an age rating of 18+.
Reading the three together (interpretation)
First, all three are documents written "before the action." The "stop button" of NO.39 is pressed after the action, and AT&T's "accuracy in the mid-to-high 90s" in NO.49 is a threshold measured while acting. These three come before that — before the agent arrives (poppy.json), before the agent is run (registry), before the answer goes out (citation, detection, notice). The order moving from after to before looks like the direction of the second half of this year (interpretation). The common field across the three documents is "actions that cannot be undone" — Poppy marks a scope `irreversible`, Uber speaks of an evidence chain that "ties privileged actions back to what the user actually requested," and Apple hands concerning questions to a person (a provider). In NO.39 we wrote that "irreversible actions belong outside the agent's hands from the start"; that sentence has become a field name in a protocol.
Second, "blocking" and "opening" are written in the same file. In poppy.json, "do not browse our website" (`web.disabled`) and "come in through the API" (`apis`) sit side by side. It means United's ban (NO.45) and Shopify's opening can become sentences of the same form, so the customer's agent reads not "blocked" but "this door is closed and that one is open" (interpretation). Three cautions when reading, though — the spec is "1.0 but a draft" with breaking changes announced, the license is stated as Apache License 2.0 but no governing entity is named, and the question CMSWire posed, "which company's agent guards the store's door" (Sierra itself sells customer-service agents), is open. "More than 100" is the count of design partners while those that have "turned it on" are three, so when relaying the numbers we keep two separate columns (NO.21).
Third, Uber and Apple are not "documents for removing people" but "documents that write the conditions for handing over to a person." Uber's registry looks like paperwork for granting autonomy, but the core is "block an action that does not fit the context just before execution" and "run new rules first in a shadow mode that only records." Apple has 200,000 people rate answers while first writing "not diagnosis or treatment" and "concerning questions go to a provider." In NO.03 we described how we made our ingredient-label app not say "safe"; Apple's document is the large edition of that — the differences are that every answer carries a citation link and the person rating the answer is the user (interpretation). The absence of figures is shared too — Uber did not give a false-block rate, Apple a duration, Sierra an actual session count. Not one number in this article was counted outside the company.
If it were us — three things for a team running commerce, apps and an AI influencer to do this week (practice)
- Fill in `/.well-known/poppy.json` for our commerce domain first "without publishing it." Moving NO.43's "three lines of rules for incoming agents" and NO.45's "why" column into this file's fields shows the blanks at once — `organization`, which doors to open (`web` only or `apis` too), whether to turn on `web.disabled`, whether to mark orders and payments `irreversible` and shipping addresses `personal_data` in `scope_hints`. Because the spec is a draft, publishing comes after the CEO's decision, and we write "1.0 = draft, breaking changes possible" at the top of the table. If any external channel carrying our products runs on Shopify, that store already has the file auto-published, so we read once what is written there.
- Write our own agent registry on one page with Uber's seven columns. Lia script generation, ingredient-label judgment (NO.03), commerce inquiry replies, and this very session that collects and writes AP News — for each, identity, owner, persona, permitted intents, constraints, dependencies (which MCP servers and tools it uses) and risk tier. It goes next to the "threshold N% for removing the human" column of NO.49. When adding a new constraint, run it like Uber in shadow mode for two weeks (record only, no blocking), then look at the false-block and false-allow counts before switching it on.
- Attach Apple's three lines to the answers of SafeList (our ingredient-label app). ① an evidence link on every answer (for us, the MFDS notices and the raw-material database are the citations), ② a list of "concerning questions" (allergic reactions, interactions with medicines, consumption by infants and the like get "ask a professional" instead of an answer), ③ whether the terms actually contain the sentence "user questions are not used to train models" and a retention period as a number (like the "24 hours" TechCrunch reported). And one rating button under each answer ("correct / not sure / wrong") — not 200,000 people, but we start counting from 20.
What we did not pick this time
The resignation of Nvidia CEO Jensen Huang and reports of his possible return, and the share-price, supply-chain and export-control coverage that followed, are a personal matter and investment territory, so we do not write about them. Coverage of Anthropic's IPO (Bloomberg) is investment territory and involves a conflict of interest, so we do not use it. The Federal Reserve survey's "87% cite AI-related risk" (Reuters) is left as a headline for the same reason. Disney's AI expansion (Bloomberg) we could not open. Meituan's LongCat-2.5 (1.6 trillion parameters, open weights, The Decoder) and Edge0-145B lack an official model card we could confirm, so they are candidates for Saturday's weekly roundup. The FAA's nationwide rollout of the air-traffic assistant SMART (Gizmodo, "human controllers keep final authority") is a "human oversight" follow-up candidate, but we did not open the primary material. "White House to ease regulatory liability for one year for voluntary reporting of AI security incidents" (AI Times, seen on promppy) — we could not obtain the original URL. OpenAI's new science blog and the Standard Model paper written with GPT-6 Astra are left as headlines because openai.com does not open from our side. Anthropic's model documentation page opened this time as an old version showing only Claude 4.x models, so we held off citing it (different from the Fable 5.1 table confirmed on November 14). Anthropic's newsroom had no new post on November 16 either, and OpenAI's news list did not open again.
What we could not confirm in this story
- The Poppy spec site's footer states Apache License 2.0 (checked directly on November 18), but no standards body or governing entity is named. "More than 100" is the number of design partners (company figure) and those actually live are three: Mercury, Synchrony and Turkish Airlines. Muse's "more than 10 million" users is TNW alone. We did not look for Shopify's own announcement page and relied on the sentence in Sierra's post about "automatic publication for every store."
- The number of sessions or orders that have actually passed through Poppy appears nowhere. Payment is written only as a "candidate extension."
- Uber's post has no figures (false-block rate, latency, cost, agent count) and no open-source statement. We could not confirm the relationship between the August ADR release report (Techstrong alone) and the November post. Nor does the post say which model company's models Uber uses.
- The duration of Apple's study is not in the Newsroom post, and "more than 12 months," "anonymized then to a third-party cloud," and "OpenEvidence deletes after 24 hours" are TechCrunch alone. The device requirement differs between TechCrunch ("iPhone 15 or later") and Apple ("iPhone 15 Pro or later"); we followed Apple. The "incomplete or unclear responses" sentence is iGeeksBlog quoting Apple's post; we could not find it in the Newsroom text.
- OpenEvidence's 760,000 physicians and 18 million monthly consultations are company figures relayed by Wikipedia. OpenAI's "300 million health queries per week" is what TechCrunch wrote; we could not open OpenAI's original.
- We did not test or count anything in any company's announcement ourselves.
What we read alongside
We opened and checked sources of different kinds directly. They are grouped as company announcements and official documents, the specification itself, media reports, wiki, and our earlier articles.
Company announcements and official documents:
- sierra.aiReleasing Poppy 1.0 (Sierra blog, Bret Taylor and Clay Bavor, November 13)English, OAuth on MCP/OpenAPI/UCP, five weeks of testing, commerce extension co-authored with Shopify, 100+ partners, Mercury, Synchrony, Turkish Airlines, "every Shopify store automatically publishes a Poppy file," on for all Muse users
- uber.comSecuring AI Agents: Lessons From Real-World Deployment (Uber Engineering, November 6)English, AI Registry, Argus, AI Guard, Agent Guard, execution trace, replay, three-layer structure, list of difficulties, no figures
- apple.comApple launches research study to explore the use of AI for health (Apple Newsroom, November 15)English, OpenEvidence, 21+, iPhone 15 Pro+, iOS 26, up to 200,000, NEJM/JAMA/Nature citations, detection of concerning questions, no training use, not diagnosis, Desai quote
- apps.apple.comApple Health AI Study App (App Store)English, developer Apple Inc., free, iPhone only, iOS 26.0+, "does not provide medical advice, diagnosis, or treatment," privacy label
Specification:
Media reports:
- thenextweb.comSierra's Poppy 1.0: Shopify turns it on for every store, Meta's Muse looks for it first (TNW, November 13)English, Muse 10M+ (one outlet), "1.0 but a draft," Turkish Airlines and Mercury examples, Visa, Mastercard, Amazon
- thenextweb.comSierra and Meta unveil Personal Agent Protocol (TNW, October 6)English, first announcement, six partners, OAuth, customer chooses read/write, company chooses the route, payments as future work
- neoteo.comSierra's Poppy draft spells out how AI agents can access businesses (NeoTeo, October 9)English, draft, /.well-known/poppy.json, 35 more design partners, Rocket demo (Sierra's own account)
- cmswire.comGenesys joins Sierra, Meta on open standard for personal AI agents (CMSWire, October 7)English, "consumers decide access, companies set limits," governance unpublished, differing partner lists, open questions
- practicalecommerce.comMeta Muse, Shopify, Amazon and the race to own AI shopping (Practical Ecommerce, September 24)English, background: Meta added to Shopify Agentic Storefronts on 9/8, Amazon's Muse-blocking message
- infoq.comUber's Secure AI framework for autonomous agents (InfoQ, November 15)English, summary of Uber's post, reader question "for agents you don't build"
- techstrong.aiUber open-sources ADR to catch agentic blunders (Techstrong.ai, August 10)English, separate August tool (company figures, one outlet) — relationship to the November post unconfirmed
- techcrunch.comApple launches new app to study how people use AI for health (TechCrunch, November 15)English, 12+ months, anonymization, third-party cloud, 24-hour deletion (one outlet), OpenAI 300M per week (relayed)
- business-standard.comApple launches app to study how people use AI for health queries (Business Standard, November 16)English, iPhone 15 Pro+, US only, no duration or compensation
- igeeksblog.comApple's AI health study is open to up to 200,000 U.S. iPhone users (iGeeksBlog, November 16)English, "incomplete or unclear responses" quote (one outlet), duration undisclosed, privacy label
Wiki:
Our earlier articles:
- apholdings.krA broadcast with no host and no camera is now on the regular schedule (AP News NO.49)
- apholdings.krWeekly AI news — OpenAI handed agents a wallet, and United left "the customer's agent" outside the door (AP News NO.47)
- apholdings.krBlock the customer's agent, let the stores in, give the model to everyone (AP News NO.45)
- apholdings.krCustomers cannot bring their own agent (AP News NO.43)
- apholdings.kr"Agents may exceed their authority" has started appearing in filings (AP News NO.39)
- apholdings.krWhen a verification code reaches an agent's hands (AP News NO.36)
- apholdings.krAgents are "untrusted insiders" (AP News NO.29)
- apholdings.krCheck the counting method before comparing numbers (AP News NO.21)
- apholdings.krWhen AI reads the ingredient label for you (AP News NO.03)