# AI agents get a front door and a wallet: Poppy 1.0, ChatGPT Wallet and Uber's intent checks

In one week, personal AI agents gained a standard way to arrive at a company's website, a wallet that pays without exposing card numbers, and a published model for checking that their actions match what the user asked for. Here is how the layers fit.

A personal AI agent that books a flight or buys a jacket for you has to solve three problems. It has to find out what a company allows agents to do. It has to prove who it is and what you let it do. And it has to pay without handing your card number to a language model. In the second week of November 2026, each of those problems got a concrete, public answer.

Sierra released version 1.0 of the Personal Agent Protocol, called Poppy, on 13 November 2026. Shopify switched it on for every merchant. According to press coverage, OpenAI launched ChatGPT Wallet the same day, so its agents can buy things using transaction-specific payment tokens. And Uber published how it checks that an agent's proposed action fits the user's request, not just the agent's permissions. Meanwhile United Airlines moved the other way and said it will restrict unauthorized third-party AI agents.

An earlier piece, [governing AI agent access](/articles/governing-ai-agent-access), covered scoped credentials in general. This one explains the new protocol and payment layer.

## What happened

### Sierra releases Poppy 1.0, and Shopify turns it on for every store (13 November 2026)

Sierra's blog post, signed by Bret Taylor and Clay Bavor, describes Poppy as a standard that "enables personal agents to interact with companies through business agents, websites, and APIs." Sierra says version 1.0 reflects five weeks of design partner testing against real websites, APIs and agents. Sierra and Meta first unveiled Poppy in October, according to The Next Web.

Sierra lists these announcements:

- Shopify enabled Poppy for every merchant on its platform. Every Shopify store automatically publishes a Poppy file, and Sierra says every Shopify store supports UCP by default.
- Poppy is on for all users of Muse, Meta's AI agent, which checks for Poppy first on every site it visits. The Next Web reports that Muse is used by more than 10 million people.
- Mercury, Synchrony and Turkish Airlines are going live as design partners.
- 60 new design partners joined, bringing the total to more than 100, Sierra says. Named partners include Delta Air Lines, Intuit, Salesforce, Stripe, Walmart and Zoom.

At Mercury, agents can check balances and start payments, which go through the account's approval policy. At Turkish Airlines, agents can check fares and find business-class upgrades, which the customer approves before paying. The Next Web notes that the specification is still labeled a draft and may change in ways that break existing setups.

### OpenAI launches ChatGPT Wallet for agent purchases (13 November 2026)

OpenAI's own announcement page could not be fetched for this article, so the details here come from press coverage. Dataconomy reports that OpenAI launched ChatGPT Wallet on 13 November. It lets OpenAI's "dots" agents, which Yahoo Finance describes as always-on agents that work in the background, complete purchases. The launch partners are American Express, Mastercard and Visa.

You add a card once. The agent proposes a purchase, and a checkout screen shows the merchant, items, total and card. You approve each purchase with Face ID or a fingerprint. The card network then supplies a transaction-specific token with its own card number, security code and expiry date. Yahoo Finance reports that the payment completes without the wallet sharing your real card details with the merchant or the agent.

Availability is narrow. Both outlets report that it starts with eligible ChatGPT Pro subscribers in the United States who are 18 or older and hold eligible US-issued cards. OpenAI plans to allow pre-approved purchases later, with limits on spending, product categories or time, according to Dataconomy.

### Uber publishes how it checks agent actions against user intent (6 November 2026, covered by InfoQ on 15 November)

Uber's engineering blog post, "Securing AI Agents: Lessons From Real-World Deployment," describes a framework InfoQ summarized on 15 November 2026. The problem, in Uber's words: identity and authorization systems "can't determine whether an agent's proposed action is appropriate in context." Uber's answer is a layer it calls behavioral authorization.

### United Airlines restricts unauthorized booking agents (11 November 2026)

Skift reports that United will restrict "unauthorized, third-party personal AI agents or assistants," on its own platforms and through travel agencies and aggregators. United cited exposed account credentials, unintended purchases and possible conflicts with FAA and DOT rules. CEO Scott Kirby told The Wall Street Journal: "I don't want to create another middleman."

## Layer one: the front door (discovery and identity)

Think of a large hotel. A guest who walks in needs to know where the front desk is, and the desk needs to know who the guest is and which rooms their key card should open. Poppy is that front desk for agents.

### The discovery file

A company publishes a file called `poppy.json` at `https://{domain}/.well-known/poppy.json`. An agent visiting a site can check this address before doing anything else. The file tells the agent:

- **Who the company is.** An `organization` block with name and domain. The spec requires `organization.domain` to match the host that served the file, so a file cannot speak for a domain that did not serve it.
- **How to sign in.** An `auth` block names the OAuth issuer and describes the permissions, called scopes, that the company offers.
- **Where to do work.** The file must list at least one of three routes. `agent.protocols` points to the company's own business agent for conversation. `apis` lists machine interfaces, each typed as `openapi` or `mcp`. `web` describes how to hand a browser session to the user.
- **What agents may not do.** A `tasks` endpoint can return `excluded_tasks`. The spec's own example is "Send a wire transfer."

Poppy does not invent a new way to place orders. It points to interfaces the company already runs, such as an [MCP](/tools/mcp) server or an OpenAPI-described API.

### How the agent says who it is

The agent's `client_id` is an HTTPS address. It returns a metadata document that points to the agent's public keys. The agent proves its identity by signing assertions with those keys, using the `private_key_jwt` method. The spec requires companies to accept the ES256 and PS384 signature algorithms.

Each request to a company's APIs or conversations also carries a `Poppy-Session` header. It holds a short-lived [JSON Web Token](/tools/jwt) that names the session, the user, the agent, the HTTP method and the target URL. These tokens are signed by delegated keys that expire within 24 hours and serve one user each, which limits the damage of a leaked key.

This is the same idea behind [agent identity and delegation](/tools/agent-identity-and-delegation): the agent proves who built it, and separately proves on whose behalf it acts right now.

## Layer two: the key card (delegated permissions)

Identity says who the agent is. Delegation says what you allowed it to do. Poppy uses standard [OAuth 2.0](/tools/oauth-2-0) for this, with the agent acting as a confidential OAuth client.

There are two sign-in paths:

1. **Direct Sign-In.** The preferred path. It uses the authorization code grant with PKCE. The request must include `prompt=consent`, so the company shows you a consent screen, unless the redirect only returns to the device that started it.
2. **Device Sign-In.** Based on RFC 8628. The agent shows a link and a code. The spec says opening the link must not approve the request by itself. You still have to click approve.

The spec is blunt about who decides: sign-in requires the user's own approval of the scopes, and the agent must not sign in on the user's behalf.

### Scopes with risk labels

Scopes are defined by each company, for example `orders:read`, `orders:write` and `addresses`. Poppy adds an `auth.scope_hints` field that labels each scope:

- `changes` says whether the scope allows no changes, reversible changes or irreversible changes. The spec says an agent must treat any unknown value as irreversible.
- `personal_data` is a true or false flag.

These labels let an agent request minimal permissions and warn you before an irreversible one. Signing out means revoking the agent's refresh tokens under RFC 7009. For how scoped OAuth works for tool calls more broadly, see [OAuth for tool access](/tools/oauth-for-tool-access).

### Secrets stay out of the model

Poppy's security section adds two rules. First, content from companies is data, not instructions. That is a direct defense against [prompt injection](/tools/prompt-injection-defense), where a web page hides commands for the agent. Second, access tokens, refresh tokens and session tokens are kept out of the model's context. The language model plans the task. Separate code holds the keys.

## Layer three: the till (checkout and payment)

Poppy handles arrival and sign-in. Sierra's post states the split plainly: "UCP handles the shopping itself, like finding products, filling a cart, and checking out." Poppy's commerce extension, which Sierra says Shopify co-authored, connects to UCP.

### How UCP works

The Universal Commerce Protocol is an open standard for interoperability between commerce entities, according to its documentation. It has its own discovery file: businesses publish a profile at `/.well-known/ucp`. That profile lists capabilities with reverse-domain names, such as `dev.ucp.shopping.checkout`. The agent and the business compare their lists and use the highest version both support. UCP runs over REST, MCP or [A2A](/tools/a2a), so the same checkout can be reached by an API client or another agent.

### Why the agent never sees your card

The UCP spec says the platform "SHOULD NOT touch raw financial credentials." Platforms send opaque values such as "tokens (such as network tokens), encrypted payloads, or mandates, not raw PANs." A PAN is the long number on your card. Businesses "MUST NOT echo credentials back in responses."

Think of a valet key. It starts the car but does not open the glovebox. A payment token works the same way: it can pay, but it is not your real card number. The ChatGPT Wallet token, as reported, is tied to one transaction, so a leak exposes far less.

### Where a human steps in

UCP builds approval into the checkout flow rather than leaving it to the agent's judgment. In UCP, if the bank requires extra authentication, the business returns `requires_escalation`, and the platform must open a `continue_url` so the buyer can complete the bank check. Required notices marked as disclosures must be shown to the buyer and cannot be hidden. For fully autonomous purchases, UCP describes an AP2 mandates extension. There the user's key signs the authorization on a non-agentic surface, which gives the business proof that the user approved that specific transaction.

| Layer | Problem it solves | Mechanism in these releases | Who sees what |
|---|---|---|---|
| Discovery | Where can an agent act here? | `poppy.json` and `/.well-known/ucp` | Public files, readable by any agent |
| Identity | Which agent is this? | Signed `client_id` metadata, `Poppy-Session` JWT | Company sees agent identity and session |
| Delegation | What did the user allow? | OAuth with consent, scope hints | User approves scopes; tokens kept out of the model |
| Checkout | What is being bought? | UCP capabilities over REST, MCP or A2A | Business sees cart, not raw card data |
| Payment | How is it paid? | Network tokens, wallets, AP2 mandates | Agent and merchant get a token, not the card number |
| Approval | Did the person agree? | Face ID or fingerprint, `continue_url`, account approval policies | User confirms; business keeps the final say |
| Intent check | Does this action fit the request? | Uber's AI Guard and execution trace | Operator sees verdicts and an audit trail |

## Layer four: the supervisor (intent alignment and audit)

A permission check answers "may this agent do this?" It does not answer "should it, right now, for this request?" An agent with a valid `orders:write` scope can still be tricked into ordering the wrong thing. Uber's framework targets that gap.

Uber describes three stages over an agent's life:

1. **Provisioning.** Every agent is recorded in the Uber AI Registry, with its identity, owner, persona, permitted intents, constraints, dependencies and risk tier. A tool called Argus analyzes the agent's code, prompts, configuration and tool-call paths. It also drafts the persona and intents so developers review them instead of writing them from scratch.
2. **Runtime enforcement.** Uber uses the classic policy pattern. The registry is the policy administration point. AI Guard is the policy decision point. Agent Guard hooks, placed at agent and tool boundaries, are the enforcement points. Before an action runs, the hook asks AI Guard for a verdict and applies it. Uber says AI Guard uses small, specialized models hosted locally that run in parallel, and may use in-memory vector search for semantic matching.
3. **Continuous assurance.** A unified execution trace links agent actions, tool calls and guardrail decisions for troubleshooting and auditing. Uber says sensitive data is minimized or redacted, and retention periods apply. Recorded interactions can be replayed against new policies before those policies are enforced.

Uber says that moving from generic controls to agent-specific policies made enforcement "more contextual and precise," as InfoQ quotes. To handle false positives, Uber used shadow-mode deployment, traffic replay and per-agent calibration.

For a shopping agent, a guard would compare the cart with your request before checkout. "Buy the blue jacket in medium" should not end in two red jackets in large. The [guardrails](/tools/guardrails) and [tracing](/tools/tracing) tools on this site cover the building blocks.

## What it means in practice

**For merchants, the agent becomes a named visitor.** Instead of guessing which visitors are bots, a store can publish what agents may do, require sign-in for account actions and mark risky scopes. Shopify merchants get this by default, according to Sierra.

**For users, approval moves to the moment of payment.** The agent searches and builds the cart. The person still says yes, through a biometric check or an account approval policy.

**For security teams, the card number leaves the blast radius.** If the agent and merchant only see a transaction-specific token, a compromised agent cannot leak a reusable card number. It can still buy the wrong thing with a valid token, which is why intent checks matter.

**The stack reuses existing standards.** Poppy builds on OAuth, JWT, MCP and OpenAPI. UCP runs over REST, MCP and A2A. A team that already runs an MCP server and an OAuth server is most of the way there. For broader context, see [agentic commerce](/tools/agentic-commerce).

## What to do now

1. **Check your discovery files.** On Shopify, see what your store now publishes at `/.well-known/poppy.json` and `/.well-known/ucp`. On your own stack, read the Poppy 1.0 spec and the UCP overview first.
2. **Inventory agent-facing interfaces.** List the APIs and MCP servers you expose and the scopes they need.
3. **Label every scope by risk.** Mark each as no-change, reversible or irreversible, and flag personal data.
4. **Keep tokens out of model context.** Hold OAuth and session tokens in the runtime, not in prompts. Treat fetched content as data.
5. **Use tokenized payment only.** Never let an agent handle a raw card number. Use network tokens, wallet tokens or signed mandates.
6. **Require approval before irreversible steps.** Show merchant, items, total and payment method on one screen. See [human-in-the-loop](/tools/human-in-the-loop) patterns.
7. **Add an intent check before tool calls.** Compare each action with the user's request. Start in shadow mode, as Uber did, and tune per agent.
8. **Log one trace per task.** Link the request, each tool call, each guard verdict and each approval. Redact sensitive fields.
9. **Decide your stance on third-party agents.** Write down which tasks you allow and which you exclude.

## Risks and open questions

**The specification may still change.** The Next Web reports that Poppy is still labeled a draft and could change in breaking ways. Early adopters should expect rework.

**Domain identity is only as strong as the domain.** Poppy's own security section notes that domain identity is not verified beyond what the domain claims. A convincing lookalike domain can still publish a valid-looking file.

**Valid tokens can still buy the wrong thing.** Tokenization protects the card number, not the decision. A hijacked agent could propose a bad purchase, and a tired user could approve it.

**Approval fatigue is real.** Once pre-approved purchases with limits arrive, the intent check carries more of the load.

**Some businesses will refuse agents.** InsideFlyer reports United's rule covers even assistants that travelers supervise or approve. Coverage will be patchy.

**Consumer readiness is unclear.** James E. Lee of the Identity Theft Resource Center told Yahoo Finance that agent payments are "a frontier technology and is probably not appropriate for most people today."

**Watch next:** whether Poppy drops its draft label, whether ChatGPT Wallet expands beyond Pro subscribers in the United States, whether other airlines follow United or Turkish Airlines, and whether intent-checking layers like Uber's appear as standard parts of agent frameworks rather than in-house builds.

## Frequently asked questions

### What is the Personal Agent Protocol (Poppy)?

Poppy is a standard from Sierra that lets personal AI agents learn what a company allows, identify themselves and sign in with the user's consent. A company publishes a `poppy.json` file pointing to its business agent, APIs or web sign-in. Version 1.0 came out on 13 November 2026.

### How is Poppy different from UCP?

Poppy covers what happens when an agent arrives: discovery, identity and delegated sign-in. UCP, the Universal Commerce Protocol, covers the shopping itself, such as products, carts and checkout. Sierra says Poppy's commerce extension connects to UCP, and every Shopify store supports UCP by default.

### Does ChatGPT Wallet give my card number to the AI agent?

No, according to press reports. After you approve a purchase with Face ID or a fingerprint, the card network issues a transaction-specific token with its own card number, security code and expiry date. Yahoo Finance reports that your real card details are not shared with the agent or the merchant.

### What does it mean to check that an agent's action matches user intent?

It means asking whether an action is right for this request, not just whether the agent has permission. Uber's framework registers each agent's purpose and permitted intents. A guard then checks every proposed action against them, along with the user's request and session history, before the action runs. Actions outside those boundaries can be denied.

### Can a company block personal AI agents?

Yes. United Airlines announced on 11 November 2026 that it will restrict unauthorized third-party AI agents, according to Skift. Within Poppy, a company can also list excluded tasks.

## Summary

Personal agents now have a clearer path from arrival to payment. Sierra's Poppy 1.0 gives them a front door: a discovery file, signed identity and OAuth sign-in with user consent and risk-labeled permissions. UCP handles cart and checkout and keeps raw card numbers away from the agent. ChatGPT Wallet, as reported, adds per-purchase biometric approval and transaction-specific network tokens. Uber's framework checks whether an action fits the request and keeps an audit trail. United's restrictions show not every company will open the door.

## Sources

- [Sierra: Releasing Poppy 1.0](https://sierra.ai/blog/releasing-poppy)
- [Personal Agent Protocol: Specification 1.0](https://personalagentprotocol.org/docs/1.0/spec)
- [The Next Web: Sierra releases Poppy 1.0 as Shopify and Meta's Muse sign on](https://thenextweb.com/news/sierra-poppy-1-0-shopify-meta-muse-personal-agents)
- [UCP: Specification overview (draft)](https://ucp.dev/draft/specification/overview/)
- [UCP: Core concepts](https://ucp.dev/documentation/core-concepts/)
- [Dataconomy: OpenAI launches ChatGPT Wallet for AI agent purchases](https://dataconomy.com/2026/11/16/openai-launches-chatgpt-wallet-for-ai-agent-purchases/)
- [Yahoo Finance: OpenAI's new ChatGPT Wallet lets AI shop for you. What are the risks?](https://finance.yahoo.com/personal-finance/banking/article/openais-new-chatgpt-wallet-lets-ai-shop-for-you-what-are-the-risks-160446618.html)
- [Uber Blog: Securing AI Agents: Lessons From Real-World Deployment](https://www.uber.com/blog/securing-ai-agents/)
- [InfoQ: Uber's Secure AI Framework Governs Autonomous AI Agents across Their Lifecycle](https://www.infoq.com/news/2026/11/uber-secure-ai-framework/)
- [Skift: United Airlines Moves to Block Personal AI Agents From Booking Flights](https://skift.com/2026/11/11/united-airlines-moves-to-block-personal-ai-agents-from-booking-flights/)
- [InsideFlyer: United restricts AI agents from managing reservations](https://www.insideflyer.com/posts/united-restricts-ai-agents-from-managing-reservations/)
