Meta Partners with Sierra to Develop Personal Assistant Open Protocol
Meta and Sierra have announced a collaboration with Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart to develop the Personal Agent Protocol, which defines how personal assistants interact with merchants. The specifications will be open to any implementers, with the first version v0.1 planned for release later this month, currently still in the co-development stage.
The announcement was made by Sierra co-founders Bret Taylor and Clay Bavor. The protocol uses OAuth to establish sessions: assistants can first check inventory or return policies as guests; when accounts are involved, users log in on the merchant's page or use credentials provided to the assistant, choosing read-only or write permissions. Questions before logging in and changes after logging in count as the same visit. Merchants can decide to open their websites, interfaces based on MCP and OpenAPI, or their own customer service assistants, with warranty claims and similar matters requiring dialogue going through the latter.
The design addresses gaps on the merchant side. Currently, most personal assistants still operate like humans, opening web pages, filling out forms, and resorting to phone calls or web chats when unsuccessful, which is time-consuming and prone to failure. Direct connections can complete the same task in seconds, but brands need to know who the assistant represents, the scope of authorization, and intent. Tony Bates, Chairman and CEO of Genesys, stated that personal AI is becoming the new front door for businesses. Kevin Miller, head of payments at Stripe, mentioned that the company joined to provide merchants with a standard way to identify customer assistants.
Amazon is not on the list. About a month ago, Meta launched the personal assistant Muse, with David Singleton, Vice President of Engineering and Consumer Products, stating it has millions of users in the U.S. and is used for event registrations, gift finding, and scheduling long-delayed appointments. On September 20, Amazon cut off Muse's shopping capabilities, stating that unauthorized AI agents accessing the service violate its terms of use. Amazon claimed that Meta did not provide prior notice, and the agent did not disclose its identity while browsing, seemingly scraping and storing customer credentials. Meta countered that Muse cannot see passwords or payment methods, and credentials are stored securely.
Payments are not included in the first version but are listed for future expansion, aiming to complete purchases without disclosing credit card information. More granular permissions can restrict specific actions, and notifications can be sent to assistants during flight delays or when orders are dispatched. Shawn Malhotra, CTO of Rocket, stated that the agent technology developed with Sierra allows Muse to move from finding housing to securing financing on its platform. Mani Fazeli, Vice President of Product at Shopify, noted that the protocol aims to convert purchase intent into results that also facilitate returns and exchanges. OpenAI and Anthropic have not joined yet, with Taylor expressing hope that competitors will adopt it, stating he would be very disappointed if they do not.
The buyers are personal assistant entities that place orders, return items, and book tickets on behalf of customers, while the sellers are merchants concerned about agents accessing accounts without disclosing their identities. This is a standard competition driven by identity gaps, not actual transaction volumes. Funds and access rights will flow to merchants willing to open their websites, interfaces, or proprietary assistants, as well as to identifiable agents. Beneficiaries include platforms like Shopify, Stripe, and Walmart that want to capture agent traffic, while those under pressure are closed storefronts that refuse unidentified agents and assistants that still rely on simulated clicks to enter sites.
Stripe and Shopify had previously joined another protocol led by Visa, and now they are entering the Personal Agent Protocol. The payment expansion not included in v0.1 indicates that the first version focuses on identity and authorization, while checkout channels will still be discussed separately.
Source: Public Information
ABAB AI Insight
Bret Taylor's path is to make the interface the industry default. He integrated Slack into enterprise collaboration at Salesforce, pushed for its sale during his tenure at Twitter, then looked at model layers as Chairman of OpenAI, and now uses Sierra to create enterprise customer service agents. Sierra raised $950 million in May, with a valuation exceeding $15 billion, selling results-based customer service rather than token-based. By bringing Meta's Muse into the same framework, it connects consumer-side agents and enterprise-side agents to the same authentication system.
Money has not been invested in unified settlement first, but in identity verification. Muse pays using saved cards through Stripe's Link, while Rocket allows Muse to transition from finding housing to securing financing, with data remaining within Rocket and Redfin's housing and loan loop. Shopify wants product inquiries, purchases, and returns to be visible to merchants. Stripe contributes a standard for identifying customer agents, rather than handing card numbers to assistants. Payments being explicitly set for future expansion indicates that the first round mobilizes access rights, while settlement remains in the original channels.
This is a parallel standard to the agent protocol that Visa has already engaged Stripe and Shopify in, not a replacement. Email's SMTP and login's OAuth were written as open specifications after various parties blocked each other. The current stage is a power struggle, not an expansion completion: Amazon cut off Muse for not disclosing identity in September, and OpenAI and Anthropic have not yet joined the working group. Walmart is on the list, while Amazon is not, thus dividing retail entry into open reception and closed storefronts.
Essentially, it is a transfer of pricing power. In the past, merchants determined who could place orders based on pages, search rankings, and their own checkout buttons. Once agents can compare prices, return items, and change itineraries on behalf of customers, clicks and shelves are no longer the only entry points; identity and authorization become new charging and release points. The mechanism is that closed storefronts can block agents that do not disclose their identities, but cannot stop users from delegating tasks to assistants. Whoever defines visitor, read-only, write, and cross-channel sessions decides whether agent traffic can convert into orders.
ABAB News · Cognitive Laws
No matter how smart the agent is, if it can't get in the door, it can't get anything done.
Openness does not mean no thresholds, but rather writing the thresholds into agreements.
Once the entry changes, pricing power shifts from shelves to identity.