AI agents are moving well beyond answering questions. Increasingly, they are being designed to book travel, manage purchases, handle customer service requests, and complete other tasks on behalf of users. A major obstacle still stands in the way, though. Businesses need a reliable way to identify these agents, verify who they represent, and control what they are actually authorized to do.

Meta and AI customer service company Sierra are working to address that challenge through the Personal Agent Protocol, an open standard designed to establish how personal AI agents connect with businesses.

The initiative was announced on October 6, 2026, with participation from Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart. On October 9, Sierra published a draft specification, known as Poppy, and announced 35 additional design partners, including OpenAI, PayPal, Visa, Mastercard, Bank of America, BBVA, Chime, Plaid, Synchrony, Target, United Airlines, Wells Fargo, and Zendesk, according to Sierra's announcement.

The expansion brings together companies from several parts of the digital economy around a common question: how can an AI agent act for a customer without compromising that customer's security, privacy, or control?

Personal Agent Protocol Targets a Growing AI Infrastructure Gap

Most personal AI agents still interact with websites and applications much as a human would. They navigate pages, fill out forms, and attempt to complete tasks through existing interfaces. When that approach fails, an agent may need to fall back on a company's web chat or contact its support team instead.

These methods can be slow and unreliable, particularly when a task involves authentication, multiple systems or changes to a customer's account.

The Personal Agent Protocol aims to provide a more consistent connection between a personal agent and a business. Rather than requiring every agent to navigate every company's website independently, the protocol describes how businesses can make their services discoverable and establish authorized sessions with agents acting for customers.

The goal is to let companies choose how an agent interacts with their systems while still giving customers control over the access they grant.

For example, a customer might ask an AI agent to check whether a product is available, initiate a return, rebook a canceled flight, or help with a service request. The protocol is intended to help the agent connect through a participating company's website, APIs, or its own AI agent, depending on what that business supports.

This could reduce friction in customer interactions considerably. However, the extent of those benefits will ultimately depend on implementation, business adoption, and the capabilities of the systems involved.

Sierra Publishes Poppy Draft With 35 Additional Design Partners

The October 9 update marks a new stage for the initiative.

Sierra said it was publishing a draft of the Personal Agent Protocol, which it calls Poppy, and adding 35 design partners to help shape its development. These companies are expected to contribute feedback drawn from their respective industries and business models.

The additional partners span banking, payments, insurance, retail, travel, communications, software, and cybersecurity. Among the named participants are

  • Banking and financial services: Bank of America, BBVA, Chime, Synchrony, and Wells Fargo
  • Payments and financial infrastructure: Adyen, Mastercard, PayPal, Plaid, Venmo, and Visa
  • Retail and consumer businesses: Gap Inc., Nordstrom, and Target
  • Travel and insurance: United Airlines, Hertz, GEICO, Insurify, and Liberty Mutual
  • Technology and software: Cloudflare, ElevenLabs, Notion, Okta, Zapier, Zendesk, and 1Password

These companies are joining the design process rather than simply endorsing a completed and universally adopted standard. Their participation could help identify practical requirements for agent authentication, permissions, integration, and customer experience.

The involvement of financial institutions and payment companies is especially relevant here. If AI agents are going to manage purchases, change bookings, or interact with financial services, businesses will need a reliable way to distinguish legitimate delegated activity from unauthorized automation.

Even so, the announcement does not establish that every partner has deployed the protocol in production or that it is already supported across their services.

How the Personal Agent Protocol Works

The protocol is built around a fairly simple principle: customers decide what access their personal agents receive, while businesses define the actions they make available.

Sierra's draft describes five building blocks that govern how agents discover a company, establish a session, authenticate, maintain continuity across channels and carry out tasks.

Discovery lets agents learn how to connect

Under the draft, businesses can publish a file at a standard website address. That file tells personal agents how to interact with the company, how customers can sign in, and which capabilities the business makes available.

This creates a consistent discovery mechanism rather than requiring each agent developer to independently work out how every website operates.

Sessions let agents act on a customer's behalf.

An agent can start a session as a guest when a task does not require access to personal account information. Checking product availability or reading a returns policy, for instance, may not require a customer to sign in at all.

The draft also allows the agent to identify itself so the business can recognize that it is interacting with an automated representative acting for a customer. This distinction matters because companies need to understand whether a request comes from a person, a personal agent, or another automated system entirely.

Signing in gives customers control over authorization.

When a task requires account access, the protocol uses OAuth, an established framework for authorizing access to digital services. The customer can sign in through the company's page, and where the business permits it, the agent can then use a session token that grants the access the customer approved.

The draft also distinguishes between access levels, including read-only access and permission to make changes. This matters because the ability to retrieve information should not automatically imply permission to execute transactions or modify an account. The protocol's intended model, then, is not unrestricted access. It is delegated access within boundaries set jointly by the customer and the business.

One session can carry across multiple channels.

A customer interaction may move between a company's website, its APIs, and its customer service agent. Without a consistent session, an AI agent may have to repeat steps or lose context each time it switches between these channels.

Poppy is designed to let a session carry across supported channels instead. A question asked before signing in and an order change made afterward can form part of the same continuous customer journey. For businesses, this could make agent-led interactions easier to manage, and for customers, it could reduce the need to repeat information or restart a task from scratch.

Businesses choose how agents complete tasks

The protocol allows participating companies to offer different routes for completing a request. An agent could use the company's regular website, connect to APIs built using standards such as OpenAPI and the Model Context Protocol, or work through the company's own conversational agent.

A retailer, for example, could let an agent use its returns page while restricting changes to sensitive account details. An airline could expose booking functions through an API, allowing an agent to find alternative flights and seek customer confirmation before finalizing a booking. In every case, the business retains control over which routes and capabilities it supports.

Why OAuth and Permission Controls Matter

As AI agents gain the ability to take actions, authentication alone is no longer enough. Businesses also need to know what an agent is allowed to do once it has been recognized.

OAuth provides a foundation for delegated authorization, and the protocol's design builds on that established approach alongside technologies including JSON Web Tokens, HTTPS, OpenAPI and MCP, according to Sierra's draft.

The distinction between identity and permission is especially important here. Identifying an agent does not, by itself, prove that it should be allowed to make a purchase, change a reservation or access sensitive customer information. A practical implementation must connect the agent's identity to the customer's authorization and the business's own policies.

The proposed approach could help companies establish clearer controls over agent-led interactions. That said, security will also depend heavily on how individual businesses implement the protocol, protect session tokens, validate requests, record actions, and handle revoked permissions. The draft is a starting point for development, not evidence that every potential security challenge has already been solved.

What the Protocol Could Mean for Banks and Payment Companies

Financial services are a natural area of interest because many customer tasks involve sensitive information, financial commitments, or changes to existing accounts.

A personal agent might eventually help a customer compare insurance policies, check the status of a financial application, manage a subscription or navigate a payment related support request. For these use cases to work reliably, businesses must be able to recognize an authorized agent and determine whether its requested action falls within the customer's permission.

The participation of companies such as Adyen, Mastercard, PayPal, Plaid, Stripe, Venmo, and Visa signals genuine interest from across the payments ecosystem in shaping how personal agents interact with businesses. That participation should not, however, be confused with a completed payments rollout. Sierra has described payments extensions as a potential future development, including the possibility of letting an agent complete a purchase without sharing the customer's credit card information.

Such functionality could reduce the need for agents to handle sensitive payment credentials directly, though the announcement does not establish that this capability is already generally available through Poppy. For banks and payment providers, the longer term question is how to preserve authorization, fraud controls and accountability as software increasingly initiates actions on behalf of customers.

How Poppy Relates to Existing Agent Standards

The AI industry is developing multiple standards to help agents interact with tools, software and digital services. The Personal Agent Protocol focuses specifically on the connection between a personal agent and a business, meaning how the agent identifies itself, obtains customer authorization and interacts with the company across supported channels.

Sierra says the protocol is designed to complement existing technologies, including OpenAPI and MCP, rather than requiring businesses to replace their current systems. The draft also distinguishes Poppy from the Universal Commerce Protocol, which focuses on shopping related experiences such as product discovery, checkout and order management. Poppy, by contrast, is intended to address a broader set of agent to business interactions, including tasks that go well beyond shopping.

This distinction matters because the agent ecosystem does not face a single technical problem to solve. Agents need ways to discover capabilities, communicate with other systems, authenticate, obtain permission and complete transactions, and different standards may address different parts of that process. Their practical value will ultimately depend on interoperability, implementation quality and the willingness of developers and businesses to support them.

The Business Challenge: Giving Agents Access Without Losing Control

For companies, supporting personal AI agents creates both an opportunity and a governance challenge at the same time.

Agents could help customers resolve problems faster, complete routine tasks and access services outside traditional customer service channels. Businesses may also gain a more consistent way to handle automated requests as a result. At the same time, companies will need to decide which actions agents can perform, what information they can access and how to respond when an agent's request falls outside its authorization.

A protocol can define a shared framework, but businesses still need their own operational safeguards. These may include access controls, activity logs, fraud monitoring, rate limits, escalation procedures and clear ways to revoke permissions. Companies must also consider how responsibility is assigned when an agent makes a mistake, misunderstands a request or carries out an action that causes financial or operational harm.

These questions become more pressing as agents move from offering recommendations to actually making changes in external systems. The standard's development process will need to account for these real world requirements if it hopes to achieve broad adoption.

What Happens Next for Personal Agent Protocol

Sierra says the Poppy draft is intended to evolve through industry feedback. The company plans to host design workshops and publish a reference implementation to help developers get started.

The wider partner group gives the initiative access to perspectives from banking, payments, retail, travel, insurance and technology all at once. That breadth could help surface integration requirements that might otherwise be overlooked in a standard developed by a smaller group.

Potential future extensions include more detailed permissions, notifications when events such as flight delays or order shipments occur, and payment functionality that avoids sharing card details with personal agents. These remain proposed directions rather than confirmed capabilities available across participating companies today.

The next important milestones will be the evolution of the specification, the availability of implementation tools and evidence that businesses can integrate the protocol into real customer journeys. Adoption will also depend on whether developers find it practical to implement and whether customers trust agents to act within the permissions they grant.

Personal Agent Protocol Signals a Shift Toward Agent Ready Businesses

The growth of AI agents is creating an entirely new layer of digital infrastructure. In the past, businesses largely designed websites and applications around people navigating interfaces themselves. Increasingly, they must also consider software acting on behalf of those people.

Meta and Sierra's Personal Agent Protocol aims to give that interaction a shared foundation, with the customer's authorization and the company's controls placed at its center.

The addition of 35 design partners, including major names in banking, payments and technology, broadens the range of businesses that can contribute to the protocol's development. It does not yet guarantee adoption, but it does give the initiative a wider base from which to test its approach.

For financial services and commerce, the stakes extend well beyond convenience. Reliable agent identity, permission management and secure execution could become essential as AI systems take on more customer facing tasks.

The key test for Personal Agent Protocol will be whether it can turn the promise of autonomous AI into dependable, authorized actions across real business systems, all without weakening the trust and control that customers expect.