Meta and Sierra’s agent protocol gets a clearer shape, but not yet a specification

Meta and Sierra’s Personal Agent Protocol is intended to let AI agents move between businesses while giving customers and companies a say over what they can do. A new report describes how the proposed sessions and permissions could work, and what the partners plan to publish next.

Meta AI Watch analysis

What happened

The protocol is designed for personal AI agents acting for consumers across company websites, APIs and businesses’ own agents. The Paypers reports that a session could begin as a guest interaction, such as checking product availability or a returns policy, before a customer signs in and grants read-only or write access for account tasks.

The report says sessions would use OAuth and carry across channels, so an interaction before sign-in and a later order change could remain part of the same visit. Businesses could let agents use their web pages or APIs, or route them to a company agent for tasks that need conversation, such as a warranty claim. Meta and Sierra have said the protocol is being developed with partners including Genesys, Instinct, Rocket, Shopify, Stripe and Walmart.

The partners plan to publish a v0.1 specification, hold design workshops and release a reference implementation. More detailed permissions, business-to-agent notifications and payment extensions are described as potential additions, not features already available.

Why it matters

The useful detail is the proposed hand-off: an agent might browse without account access, then receive narrower permissions when a task requires them. That could make agent-led shopping and service less like handing over the keys and hoping for the best. It also puts the hard questions in view: what can an agent change, who sets those limits, and how can a customer see what happened?

A shared approach could save each business from inventing its own way to recognise agents. But until the specification arrives, the report sketches intended behaviour rather than giving developers something they can implement or customers something they can use.

Our read

This is a more practical outline than a promise that agents will simply “do commerce”. The proposed guest-to-account flow and planned permissions are worth watching because they put control at the centre of the transaction, not in a footnote after checkout.

For now, treat the mechanics as a reported plan. The v0.1 specification and reference implementation will show whether the partners can turn a sensible sketch into rules that are clear to users as well as convenient for businesses.

What to watch

• Whether the planned v0.1 specification and reference implementation are published, and when.
• How customers grant, narrow or revoke access for individual actions.
• What records customers and businesses can see after an agent acts.
• Whether the proposed notifications and payment extensions make it into later versions.

Discussion spark: Should customers grant AI agents broad account access for convenience, or should every consequential action require a separate permission?

not affiliated with, endorsed by, or operated by Meta