Meta and Sierra announce the Personal Agent Protocol, an OAuth-based open standard for personal AI agents
On October 6, 2026, Sierra and Meta announced the Personal Agent Protocol (Poppy), an open draft standard for how personal AI agents discover companies, sign users in over OAuth, and act within user-granted scopes.
What happened
Sierra and Meta announced the Personal Agent Protocol on October 6, 2026, in a post by Bret Taylor and Clay Bavor. It is an open standard for how a consumer's personal AI agent works with businesses: finding a company, signing the user in, and acting through APIs, web pages or the company's own agent. The project calls it Poppy for short.
Key points from the launch post and the specification site:
- Industry partners named at launch: Genesys, Instinct, Rocket, Shopify, Stripe and Walmart.
- Sign-in is built on OAuth. The user signs in on the company's page and chooses read-only or write access.
- Companies publish a discovery file at /.well-known/poppy.json and define their own scopes.
- The spec is Apache 2.0 and open to anyone to implement. It is not hosted by a standards body.
- Versions move quickly: 0.2 was published October 24 and 0.3 on November 4, 2026; 0.4 is an open draft. Payments, push notifications and finer-grained permissions were listed as later work.
Why it matters
Consumer agents have mostly reached accounts by logging in as the user with a shared password, a pattern banks are now warning customers against. PAP gives companies a sanctioned alternative built on the delegated-access tools CIAM teams already run: an authorization server, scopes, consent and revocable tokens. It is backed by a large platform and several large merchants at once, which is what a consumer-agent standard needs to get adopted.
Deepak's take
The protocol itself is less new than the coalition behind it. OAuth, scopes and discovery metadata are old tools. What changes is that a large consumer platform and major merchants agreed to use them for personal agents, instead of letting agents scrape and log in as the user.
The pace is a caution. Three numbered versions in five weeks means early implementers will chase changes. That is normal for a 0.x spec, but pin a version and expect to update.
The hard part for companies will be scope design. A scope a person can understand ("see my orders", "place orders under a limit") is a product decision, and it is the consent record you will rely on when an agent does something the customer did not expect.
What to do
- Consumer-facing companies: Inventory which actions you would let an agent take, and draft scopes in plain language before choosing a protocol.
- CIAM teams: Confirm your authorization server can issue agent-specific grants and revoke them without a password reset. See agent-level consent.
- Early adopters: Build against a published, frozen version, not the latest draft.
- Further reading: authorization patterns for agentic workflows and agent credential sharing.