
On October 6, 2026, at the Sierra Summit in San Francisco, Sierra (co-founded by Bret Taylor) and Meta unveiled the Personal Agent Protocol (PAP), an open standard designed to let a customer's personal AI agent talk to a business's systems. In practice: instead of a human visitor browsing a website, it is an AI agent acting on the customer's behalf that shows up, identifies itself, and requests to perform an action. For any SMB that sells online or manages remote customer relationships, this AI agent protocol is worth watching before it becomes a de facto standard.
In brief
- On October 6, 2026, Sierra and Meta announced the Personal Agent Protocol (PAP), an open, OAuth-based standard for authenticating personal AI agents with businesses.
- Six founding partners joined the initiative: Walmart, Shopify, Stripe, Genesys, Rocket, and Instinct.
- The protocol distinguishes guest access (checking stock, a return policy) from signed-in access, with read-only or read/write permissions chosen by the customer.
- A v0.1 specification is planned for late October 2026, along with design workshops and a reference implementation to come.
- A notable detail for judging the standard's reach: Amazon, OpenAI, and Anthropic are not part of it at this stage, and a PartnerCentric study cited by the Wall Street Journal found that only 3% of US adults trust an AI agent to make a purchase on their behalf.
The problem the PAP tries to solve
A personal AI agent (a customer's assistant, built into their phone or a messaging app) can already browse a website, fill out a form, or attempt a purchase. But nothing tells the business receiving that request whether it comes from an agent authorized by its owner, an unidentified bot, or an impersonation attempt. Sierra describes this lack of a shared standard as "kind of chaos" until a norm takes hold. Every network and every vendor was building its own verification, which would have multiplied the technical integrations each merchant needs just to welcome these agents properly.
The PAP sets a common layer to answer two questions: who is speaking on the customer's behalf (authentication), and what is that entity allowed to do (permissions). It does not touch payment settlement itself, which stays with existing financial networks: that is an important difference from transaction-focused initiatives such as the Know-Your-Agent framework from Visa, Mastercard, and Ant International, announced in September 2026.
2025 - early 2026
Fragmented protocols
October 6, 2026
Personal Agent Protocol announced
Late October 2026
v0.1 specification
Upcoming
Planned extensions
How the authentication works
The mechanism Sierra describes relies on OAuth, an authorization protocol already widely used on the web (it is what lets you log into a service with your Google account, for instance).
Guest session
Customer sign-in
Choosing the access level
Business-defined scope
The session stays valid across channels: a question asked before sign-in and an account change made after sign-in count as a single continuous visit, according to Sierra's description.
Why Meta and six partners are joining forces
For Meta, the stake is that its own agents (including Muse, the assistant that rolled out built-in checkout with Shopify as early as September 2026) can interact with as many businesses as possible without case-by-case negotiation. For the partners, the logic differs by line of business:
Without a shared standard
Every business builds or negotiates its own integration, agent by agent. Repeated technical costs, the risk of blocking legitimate agents out of excess caution, or conversely letting fraudulent requests through.
With the Personal Agent Protocol
A single technical integration recognizes every PAP-compatible agent. The business keeps fine-grained control over what it allows (read, write, or nothing at all) and can route complex requests to its own agent.
The limits to keep in mind before getting too excited
The launch messaging remains promotional, and several caveats are worth noting before treating the PAP as tomorrow's default standard.
Another point to watch: Amazon, OpenAI, and Anthropic are not among the founding partners announced. Bret Taylor says he hopes they will join the standard eventually, without giving a timeline. Amazon has reportedly restricted some competing agents' access to its site and taken legal action against Perplexity over data-collection concerns, which illustrates the competitive tensions around this topic. Finally, the full technical specification has not been published yet: what is described today is a framework of principles, not yet an operational standard that an SMB's technical team could implement.
What this concretely changes for an SMB
Even at the specification stage, the PAP (and similar initiatives) point to a clear direction: part of tomorrow's customer traffic will no longer come from a browser, but from an AI agent acting on a customer's behalf. A few useful reference points:
| Situation | What to anticipate |
|---|---|
| E-commerce or booking site | Check whether your platform (Shopify, Stripe, etc.) plans to support the PAP or an equivalent protocol, and decide which actions an agent may take without human validation. |
| Customer service | A "guest" AI agent should be able to check a return policy or order status without an account, which reduces support load for these simple questions. |
| Security and fraud | Write access (placing an order, changing a booking) should remain conditional on clear customer authentication, not merely on the presence of an agent claiming to be authorized. |
| Choosing tools | Favor providers (e-commerce platform, CRM, payment processor) that are participating in discussions around these standards, to avoid a costly in-house integration later. |
For now, this standard is more of a signal than an obligation. But an SMB that understands early the guest / signed-in / read or write logic will be better placed to decide, in the coming months, what it actually allows on its site when visitors are no longer humans behind a screen.
FAQ
What is a personal AI agent?
A personal AI agent is a smart assistant, built into a phone, a messaging app, or an application, that acts on a user's behalf: looking up information, filling out a form, or in some cases triggering a purchase or a booking, with its owner's authorization.
Does the Personal Agent Protocol handle payments?
Not directly. The PAP defines the agent's identification and its access permissions (read or write). Financial settlement stays with existing payment networks; a payments extension is only mentioned by Sierra as a future development.
Does an SMB need to act right now?
Not urgently: the v0.1 specification is only expected by late October 2026, and the standard does not yet include Amazon, OpenAI, or Anthropic. It is still worth tracking its evolution, especially if your business relies on an e-commerce site, a booking platform, or high-volume customer service.
How is this different from the MCP or A2A protocol?
MCP (Model Context Protocol) and A2A (Agent2Agent) organize communication between AI systems themselves (an agent using a tool, or two agents coordinating with each other). The Personal Agent Protocol sits at a different level: it governs the relationship between an end customer's AI agent and a business's system, on the authentication and permissions side. The two approaches are complementary rather than competing.
To keep up with other developments in agentic standards and what they mean for SMBs, browse our full AI watch or see how businesses like yours are already adapting their tools in our customer stories.


