Ai Agents 7 min read

Poppy 1.0: Sierra and Meta's personal agent protocol goes live on every Shopify store

Sierra released Personal Agent Protocol 1.0 on November 13, 2026. Shopify now publishes a poppy.json file for every merchant and Meta's Muse checks for one first. Here is how discovery, agent identity and scoped access work, and what the spec still leaves out.

On November 13, 2026, Sierra published Personal Agent Protocol 1.0, an open specification, nicknamed Poppy, for how a personal AI agent acting for one person deals with a company’s website, APIs and support agent. The post is signed by Sierra co-founders Bret Taylor and Clay Bavor, and the protocol is developed with Meta. Two adoption facts make the release matter: Shopify says it “has enabled Poppy for every merchant on its platform”, so every Shopify store now publishes a Poppy file automatically, and Meta’s Muse agent “automatically looks for Poppy first on every site it visits.”

That combination gives the spec something most agent standards lack on day one: a large set of sites that already serve it, and a widely used agent that already reads it. The question every company has been fighting over in public, which agents are allowed in and what they may do, now has a machine-readable answer format.

What shipped, and who is behind it

Sierra introduced the protocol on October 6 with Meta, Genesys, Instinct, Rocket, Shopify, Stripe and Walmart, published a draft three days later, and called the November 13 release 1.0. According to the project’s path to 1.0 page, version 1.0 is version 0.4 “with fixes for problems found while implementing it”, and versions 0.1 through 0.4 remain on the site.

The design partner list now runs past 100 organizations. On the agent side it names four: Instinct, Manus, Meta and OpenAI. On the company side it includes payments firms (Stripe, Adyen, PayPal, Visa, Mastercard, Affirm, Checkout.com), banks (Bank of America, Wells Fargo, BBVA, Navy Federal Credit Union, SoFi), retailers (Walmart, Target, Wayfair, Nordstrom, Gap), airlines (Delta, United, Turkish Airlines), and identity vendors (Okta, Ping Identity, Persona, Descope, WorkOS). Sierra’s release post names three companies with live, task-level support: Mercury, where an agent can check balances, review transactions and initiate payments subject to the account’s approval policy; Turkish Airlines, covering fares, flight status, stopovers, business-class upgrades and Miles&Smiles; and Synchrony, which is described as still exploring.

United’s place on that list is worth noting. On November 12, a day before the 1.0 release, United said it would block “unauthorized, third-party personal AI agents” from booking flights, a story we covered alongside OpenAI’s ChatGPT Wallet launch. Poppy is designed to be the authorized path: a way for a company to recognize an agent, know whose account it is acting on, and decide what it may do, rather than blocking all automated traffic.

How the protocol works

The 1.0 spec layers six pieces: discovery, operator recognition, sessions, account access, interaction channels, and task coverage through extensions. Almost all of it reuses existing standards.

Discovery. A company “MUST publish its poppy.json document at https://{domain}/.well-known/poppy.json” over HTTPS. The file declares the protocol version, the organization (name, domain, logo), an auth block, and at least one of three interaction channels: agent (an endpoint for the company’s own support agent), apis (OpenAPI 3.x or MCP servers), or web (browsing the site). A tasks endpoint and an extensions map cover anything else. A website-only company can publish the file with no auth section at all, which is part of why Shopify could turn it on for every store at once.

Permissions. Scopes such as orders:read, orders:write or addresses are declared in the file, and each can carry scope_hints saying whether it causes no change, a reversible change or an irreversible one, and whether it touches personal data. The user decides what to grant, and Sierra’s October post frames it simply: the customer decides “whether the agent has read-only or write access.” An agent can also start as a guest with no sign-in.

Account access. Sign-in is standard OAuth. Authorization server and resource metadata live at the usual well-known paths (RFC 8414 and RFC 9728). The preferred flow is authorization code with PKCE; device sign-in (RFC 8628) and a third “mediated” grant are also defined, and DPoP token binding is recommended.

Agent identity. Each personal agent’s client_id is an HTTPS URL that returns a client metadata document with its signing keys, so a company does not need to pre-register every agent. The agent authenticates to token endpoints with a signed JWT assertion, and companies must accept ES256 and PS384. Every request to a company’s API, MCP server or endpoint then carries a Poppy-Session header: a short-lived signed JWT naming the agent, the company, the session, the HTTP method and URL, and a user ID. That user ID is pairwise, meaning each company sees a different opaque identifier for the same person, so two companies cannot join their records on it. For the operator of the agent itself, the spec recommends Web Bot Auth signatures.

How it fits with MCP, UCP and robots.txt

Poppy is not a replacement for MCP or for commerce protocols. Sierra says it is “largely built with existing standards like MCP, OpenAPI, and UCP.” An existing MCP server can be listed in poppy.json unchanged except for accepting the session header, and Sierra’s company kit says the server “can stay as it is.”

The split with Shopify’s Universal Commerce Protocol is explicit. UCP covers shopping itself: search, cart and checkout. Poppy covers what happens before that: in Sierra’s words, “when an agent shows up at a store”, says who it is, and finds out what it is allowed to do. Shopify co-authored Poppy’s commerce extension, which makes an agent ask the user before certain actions.

The spec does not reference robots.txt. Its closest equivalent is a web.disabled flag, set to true when a company asks personal agents not to browse its website in a browser. That makes Poppy a richer signal than a crawler rule, since a company can turn off browsing while still offering an API or its own support agent, but it is also only a request: like robots.txt, it relies on agents choosing to honor it.

What it does not cover yet

The 1.0 label is softer than it sounds. The GitHub repository README still describes the site as publishing the “draft specification,” the company kit says “the spec is a draft,” and the path to 1.0 page says some fixes “broke compatibility with earlier versions.” Teams building against it should expect more changes.

Several pieces are explicitly out of scope or on the roadmap: payments (Sierra’s October post floats extensions that would let an agent pay “without sharing credit card information”), push notifications, attachments, signed approvals, self-hosted personal agents, and removing the mediated sign-in grant. Permissions are scope-level today; finer-grained, per-action permissions are listed as future work. How a company decides which agents to accept is left to each company.

Governance is light. The spec and docs are Apache 2.0, but the names “Personal Agent Protocol” and “Poppy” are excluded from that license, only collaborators can open pull requests on the repository, and there is no neutral foundation or standards body behind it yet. Sierra builds customer service agents for many of the companies on the partner list, so it sits on both sides of the protocol it authored.

What to watch

The adoption test is on the agent side. Meta and OpenAI are design partners, but only Muse has been named as reading the file by default; whether OpenAI’s dots and other agents check for poppy.json will decide whether companies bother to publish more than the minimum. Amazon, which blocks third-party shopping agents and steers customers to its own assistant, is not on the list. Watch for live API scopes from the banks and airlines on the partner roster, a payments extension that could meet efforts like ChatGPT Wallet, and whether the spec moves to a neutral home before more companies build on it.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Agents

ADK Adds Native Live Evaluation for Voice Agents

Learn how to evaluate real-time voice agents in Google ADK with audio simulation, trajectory rubrics, CLI automation, and ADK Web playback.

Ai Agents

Meta Acquires Moltbook, Bringing Viral AI Agent Network's Founders to Superintelligence Labs

Meta acquired Moltbook and hired its founders into MSL, betting on AI agent identity and directory tech after the platform's spoofing scandal.

Ai Agents

Cloudflare: 50+ AI pentest frameworks scanned its customers in three days, and traffic is up 30x since October 1

Cloudflare's November 13 report says taking ARTEX offline produced a hydra effect: hundreds of forks, 50+ framework families and a 30x rise in agentic attack traffic. Here is what the data shows, what it leaves out, and what a stolen CI/CD token looked like in one incident.

Ai Agents

Google's /teamwork results: 1.8 million lines of Lean for the Weil conjectures, a 2013 conjecture proved, and a Rust Zircon

On November 13, Google Research published what its Antigravity /teamwork multi-agent harness produced over several weeks: a proof of the Courtade-Kumar conjecture, a machine-checked formalization of the Weil conjectures, and 800,000 lines of Rust for the Zircon kernel. Here is how the harness works and what the post leaves out.

Ai Agents

ChatGPT Wallet lets OpenAI's dots pay with network tokens, no OpenAI cut at launch

OpenAI launched ChatGPT Wallet on November 13, 2026, letting its always-on dots agents buy things with scoped Visa, Mastercard and Amex tokens. Here is how the payment path works, what merchants must change, and why it starts so small.