# When the customer's agent says yes, who agreed?

> Personal Agent Protocol 1.0 lets a customer's agent sign in and talk to a brand's own agent. It settles which agent is calling and what it may do. Whether the customer agreed is left to the brand.

By [Ron Fogel](https://bazak.ai/writing/authors/ron-fogel), Co-founder & CTO · Published 2026-11-14 · Agentic commerce, Engineering

On 11 November the European Commission wrote AI shopping agents into a draft law. Two days later Sierra, which develops the Personal Agent Protocol with Meta and industry partners, published version 1.0. The protocol goes by Poppy. A customer's personal agent can now sign in to a business and talk to the business's own agent.

The specification is careful about identity and access. It leaves one question to the team that runs the agent on the business side. When the customer's agent says yes, did the customer agree?

## What Poppy settles

Poppy starts with discovery. A business publishes a file that tells a personal agent how to reach it, and the business chooses which doors to open: its website, its APIs, or its own company agent. The file has to list at least one.

Identity comes next. A personal agent's ID is a URL that returns its metadata document, and the agent proves it holds the matching key.

Access runs on OAuth. The customer signs in at the business, sees the scopes the agent asked for, and approves some, all or none of them. The business may not grant more than the agent requested. Each scope can say whether its actions change nothing, change something that can be undone, or change something that cannot. The customer can revoke access later, and a business can bind its tokens to the agent's key with DPoP.

Adoption is already wide. Sierra counts more than 100 design partners and says the protocol is built largely on existing standards such as MCP, OpenAPI and UCP. Meta's agent Muse looks for Poppy first on every site it visits. Every Shopify store now publishes a Poppy file automatically. Shopify also co-authored a commerce extension that links Poppy to UCP, lets an agent shop as a guest or signed in, makes the agent ask the customer before certain actions, and connects to UCP Ask so a shopper's agent can talk to the store's agent.

After all of that, a company agent knows which agent it is talking to, on whose account, and what the account holder allowed.

## What it leaves open

Turkish Airlines is live on Poppy. According to Sierra, a personal agent can check fares, flight status and stopovers, and pull up a booking with a reference number and a surname. Members of the airline's loyalty program can have their agent find business-class upgrades, which the customer approves before paying.

Now imagine a carrier that skips that last step. A traveler tells their agent to see whether they can sit somewhere better on Thursday's flight. The personal agent signs in with a scope that allows changes to the booking. The carrier's agent offers an upgrade for 240 euros. The personal agent answers "Confirm." The traveler has not seen the price.

Every check passed. The token was valid, the scope covered the action, and the agent was the one it claimed to be. The carrier still holds no evidence that a person agreed to pay 240 euros.

The spec saw this coming. Each conversation message carries an author field (sections 8.4 and 8.8). Its type is `agent` if the personal agent wrote the message, and `human` only when the message carries the user's own words. A summary or a paraphrase of what the user said counts as `agent`, and so does any value the receiver does not recognize. Section 8.11 closes the loop: "Sending a message does not authorize the requested account action."

So a company agent has two counterparties in one conversation. One is software acting on instructions the company cannot read. The other is the account holder, who may be following along or may be asleep. A yes from the software tells you what the software decided.

The Digital Fairness Act proposal gives the difference legal weight. Under Article 26, a consumer who starts using an agent chooses whether to confirm all contracts, some of them within a maximum price and a time limit, or none. When the agent's provider is paid for proposing a contract, or has a conflict of interest, the consumer has to confirm that contract whatever they chose. Those duties sit with the provider of the personal agent. Nothing in the article requires the personal agent to tell the brand on the other side which setting the consumer picked, or whether a payment stands behind the recommendation. The author field is the signal it does get, and `agent` never means the consumer confirmed.

## Four principles for the company agent

**Treat agent-authored messages as untrusted input.** A personal agent reads web pages, email and other companies' replies before it reaches you, and it can carry instructions it picked up on the way. The spec tells personal agents to treat content from a company as data. For conversation messages it gives companies no matching rule, so write your own.

**Never let an agent's message satisfy a human-confirmation step.** The Operations extension gives the company a tighter loop. The company refuses the action and returns its exact terms. In the basic case, "the Personal Agent asks the User, and the User approves." The extension also lets a personal agent approve on a standing instruction from the user, and it lets the company refuse that kind of approval for a given action. Either way the approval reaches the company as the personal agent's statement, and the extension says so. When the company needs to see the customer approve, the extension points outside itself: the customer confirms on the company's own website or app, signed in. Where money moves, choose between those paths on purpose.

**Keep scopes narrow and tied to actions with money at stake.** A single "manage my booking" scope puts a free seat change and a paid upgrade behind one grant. Split them, and mark the ones that cannot be undone. The protocol has a field for it.

**Record provenance.** For each message, keep who wrote it, under which scopes, and who approved which terms. Article 34 of the proposal lets a court or an authority require a trader to disclose evidence about the design, operation and impact of an online practice. A failure to disclose can count toward a finding that the practice was unfair. A log that cannot tell the agent's yes from the customer's yes will not answer that request.

## What we do at Bazak

We build agents for brands, and we design them on the assumption that the party on the other side may be another agent. We treat confirmation as the customer's alone.

## Where this leaves teams

Poppy gives a company agent a reliable answer to two questions: which agent is calling, and what it may do. Consent is the next problem, and a protocol field cannot produce it. The company agent has to ask for it from the right party, on the exact terms, and keep the record. That work belongs to the team that runs the agent.

## Sources

- [Sierra, Releasing Personal Agent Protocol 1.0, 13 November 2026](https://sierra.ai/blog/releasing-poppy)
- [Personal Agent Protocol specification (draft 1.0)](https://personalagentprotocol.org/docs/1.0/spec)
- [Personal Agent Protocol 1.0, Operations extension](https://personalagentprotocol.org/docs/1.0/extensions/operations)
- [European Commission, Proposal for a Digital Fairness Act, COM(2026) 1100](https://commission.europa.eu/document/download/0575d59f-554f-4a9e-b54b-e6b717175809_en?filename=Proposal+for+a+Digital+Fairness+Act.pdf)

---

[Bazak](https://bazak.ai) · [Get a demo](https://bazak.ai/demo) · [All pages](https://bazak.ai/llms.txt)
