Learn  /  What's new

The Personal Agent Protocol explained: how AI agents will log in, shop and act for you

Your AI agent will soon check stock, change orders and book appointments for you. A new open standard decides how it signs in and what it may do. Here is how it fits with MCP and A2A.

What's new8 min readUpdated 21 October 2026
AIAI agentsWhat's new
All guides and cheat sheets

Personal AI agents, assistants that act on your behalf rather than just chat, are arriving fast. OpenAI launched always-on "Dots" at DevDay, and Meta's Muse agent can already pay with saved cards. But when your agent visits a shop or a bank, how does the business know who it is, what it is allowed to do, and that you approved it? On 6 October, Sierra and Meta proposed an answer: the Personal Agent Protocol, an open standard built on OAuth and backed by Walmart, Shopify, Stripe and others.

v0.1SpecificationOctober 2026
8Named backersincluding Walmart and Stripe
OAuthBuilt onread-only or write scopes

How it works

PAP describes the conversation between a customer's agent and a business, from an anonymous first question to an authorised action:

A personal agent changing an order with PAP
You"Move my delivery to Friday"
Your agente.g. Meta Muse or a Dot
Guest accesschecks stock or policies, no sign-in
OAuth sign-inyou grant read or write access
Businesswebsite, API or its own agent
Guest access
An agent can ask public questions, such as stock levels or a returns policy, before anyone signs in.
Scoped permissions
After the customer signs in, they choose whether the agent gets read-only or write access.
One session across channels
Questions asked before sign-in and actions taken after belong to the same visit.
Business choice of interface
A company can let agents in through its website, its APIs (for example via MCP or OpenAPI), or its own customer-service agent.
Visibility
Businesses can see what agents are doing on their sites, rather than guessing from unusual traffic.

PAP, MCP, A2A: which protocol does what?

A year ago "agent protocol" mostly meant MCP. There are now several, and they solve different problems:

MCP

Agent ↔ tools
Started by
Anthropic, 2024
Connects
an AI app to tools and data
  • The standard way to plug databases, files and APIs into AI assistants
  • Supported by every major AI platform

Use it to give an agent access to your data.

A2A

Agent ↔ agent
Started by
Google, 2025
Connects
agents to other agents
  • Lets agents from different vendors delegate tasks
  • Discovery through published agent cards

Use it when your agents need to work with someone else's.

PAP

Personal agent ↔ business
Started by
Sierra and Meta, 2026
Connects
a customer's agent to a company
  • Identity and permissions for consumer agents, built on OAuth
  • Version 0.1; payments not included yet

The one that matters for customer-facing businesses.

The payments question

Payments are listed as a future extension of PAP. In the meantime, card networks and payment companies have their own standards. Visa runs the Trusted Agent Protocol, which Stripe, Shopify, Microsoft and Worldpay have joined, so some of the same companies back both. Expect the two to meet: PAP for identity and permissions, a payment protocol for the money.

Who is behind it

Sierra
Co-author. Builds customer-service agents for large companies; announced the protocol.
Meta
Co-author. Its Muse personal agent is one of the first consumer agents that will use it.
Walmart and Shopify
Retail partners: one of the largest retailers and the platform behind millions of online shops.
Stripe
Payments partner, also a member of Visa's Trusted Agent Protocol.
Genesys
Contact-centre software, where agent-to-agent customer service will happen.
Rocket and Instinct
Further launch partners named in the announcement.

What it means for data and analytics teams

When agents shop, book and complain on customers' behalf, a growing share of "customer" activity comes from software. On this site, analytics already flag visits from AI assistants such as Google's NotebookLM reading articles. For businesses, that shift has practical consequences:

  1. 1

    Identify agent traffic

    Tag sessions that arrive through PAP, MCP or known agent user agents so they do not distort conversion rates, funnels and A/B tests.

  2. 2

    Measure agents as a channel

    Report agent-led orders, changes and support contacts separately, just like mobile or email.

  3. 3

    Expose clean data

    Agents work best with clear APIs and well-described data: product catalogues, stock, order status and policies. The same semantic clarity helps internal AI too.

  4. 4

    Log permissions

    Record which agent acted, under which scope and with whose approval. It is the audit trail compliance teams will ask for.

Check your understanding

1/3What is the main job of the Personal Agent Protocol?

2/3Which technology is PAP built on?

3/3Are payments part of PAP version 0.1?

Agents are only as good as the context you give them. See how the data platforms are preparing in Databricks: October 2026 update, and which models power these agents in The best AI models in October 2026.

Sources: The Next Web on the Personal Agent Protocol (6 October 2026) and Sierra's announcement by Bret Taylor and Clay Bavor; AI Weekly; reporting on Visa's Trusted Agent Protocol. The specification is at version 0.1 and may change.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 21 October 2026. About me

Found this useful?

New guides and "what's new" explainers on data and AI land here regularly. Follow along on LinkedIn, or share this one with a colleague.