Meta and Sierra’s Personal Agent Protocol would give AI agents a front door to businesses

What was announced?

Sierra, the customer-service AI company led by Bret Taylor and Clay Bavor, published a post on 6 October 2026 introducing Personal Agent Protocol. Sierra says Meta and Sierra are developing it together with industry partners at Genesys, Instinct, Rocket, Shopify, Stripe and Walmart, and that it is “open for anyone to implement.”

The protocol addresses a specific gap. Personal agents that book flights or shop for insurance today often work the way people do: loading pages, filling forms, and sometimes calling a support line or opening a web chat. That kind of computer-use agent can be slow and brittle. Sierra argues that a direct, authorized connection could finish the same task securely in seconds.

The technical details here come from Sierra’s announcement of what it describes as a joint project with Meta. In an interview with CNBC, Taylor, who is also chairman of OpenAI, said OpenAI and Anthropic are not part of the effort for now but that he expects them to participate.

How would Personal Agent Protocol work?

Sierra summarizes the principle simply: consumers decide what access to give their personal agents, and companies set parameters for what those agents can do. The flow it describes has four steps.

The design leans on existing standards instead of inventing new ones. OAuth already handles delegated access on much of the web, and Sierra names MCP and OpenAPI as the expected interface standards for direct API access.

1. Discovery: the protocol starts on the company’s website, where a personal agent can learn what the company offers and how to reach it.
2. Session: the agent starts a session for its user, possibly as a guest, which may be enough to check product availability or a returns policy.
3. Sign-in: when a task needs account access, the customer signs in on the company’s page or uses credentials already set up with the agent, and chooses read-only or write access.
4. Execution: the agent completes the task through the company’s website, its APIs, or the company’s own agent for conversational work such as a warranty claim.

How does it relate to MCP and other agent standards?

Personal Agent Protocol sits above interface standards rather than replacing them. The Model Context Protocol defines how an AI application connects to tools and data. Personal Agent Protocol, as described, covers what comes before and around that: identifying a business, establishing whose behalf an agent is acting on, and carrying one authorized session across a website, an API and a company agent.

Sierra lists several possible extensions, but only as possibilities: more detailed permissions for specific actions, push notifications such as a flight delay or shipped order, and payment extensions so an agent could complete a purchase without sharing card numbers. None of these is described as part of the initial specification.

Who benefits, and what changes for businesses?

Sierra frames the protocol as balancing three groups. Consumers want tasks done correctly the first time by an agent acting in their interest. Brands want to know when an agent is acting for a customer and to decide what it may do. Agent builders want one consistent way to work with participating companies instead of a custom integration for each.

For a business, the practical appeal is visibility. A scraping agent looks like ordinary traffic; an agent arriving through a declared protocol session can be identified, scoped and audited. Genesys chairman and CEO Tony Bates is quoted in the announcement saying brands need a trusted way to know who an agent represents, what it is authorized to do, and its intent.

The partner list spans commerce platforms, payments, retail and contact-centre software, which suggests the first use cases will be shopping, account servicing and customer support. That is an inference from the partners named, not a published roadmap.

What are the risks and open questions?

The biggest caveat is that there is nothing to evaluate yet. Without a published specification, it is not possible to judge how scopes are defined, how agents are identified, how revocation works, or how disputes over an agent’s actions would be handled. Delegating write access to an agent also raises the stakes of prompt injection: an agent that can change an order or a booking needs strong approval gates for consequential actions.

Adoption is the other open question. Open standards succeed when competing agent builders and large merchants implement them. According to CNBC, OpenAI and Anthropic have not joined yet, and the partners are listed as collaborators on development, not as companies that have shipped support.

What happens next?

Sierra says it plans to publish the v0.1 specification later in October 2026, host design workshops with interested parties, and publish a reference implementation to help developers start. It invites other companies and agent builders to join; Instinct is named as joining the effort.

Developers and customer-experience teams who want to influence the design should watch for the v0.1 draft and workshop announcements. Until then, the safest assumption is that existing integrations, such as public APIs and MCP servers, remain the way agents reach businesses.