Identity · CA
A New Protocol Lets Businesses Verify Which Human a Personal Agent Speaks For
Decagon, the AI customer service company behind the personal agent "Instinct," open-sourced a new standard this week called the Personal Agent Consent & Trust Protocol, or PACT. It tackles a problem that's been quietly building as personal agents start booking travel, managing purchases, and resolving disputes on someone's behalf: when that agent shows up at a business's AI support system, how does the business know who it's really acting for, and what it's actually allowed to do?
Decagon describes the shift directly: personal agents like Muse, Instinct, and dots are starting to handle real transactions, and as they contact businesses, they increasingly run into AI agents working the other side of the conversation. PACT is meant to give that interaction explicit, verifiable customer permissions rather than leaving it to implicit trust or screen-scraping.
The mechanism is layered on existing standards rather than invented from scratch. PACT is built on Google's Agent2Agent protocol and OAuth 2.0, and it lets businesses verify whom a personal agent represents and what that customer has authorized it to do. In practice, the personal agent discovers the business's A2A Agent Card — the same capability-advertising document A2A already defines — which lays out the business's endpoint, authentication requirements, and available permissions. Each business then defines its own scopes, things like orders:read or orders:cancel, with human-readable descriptions the personal agent uses to figure out which permissions it actually needs for the task at hand.
What makes this notable isn't the cryptography — it's the scoping model. Rather than a binary "is this an AI or a human" check, PACT pushes toward fine-grained, per-business permission grants that travel with the agent into each new interaction. That's a meaningfully different shape of problem than the authentication question Meta, Walmart, Stripe, and Sierra raised days earlier with their own personal agent protocol effort — that work focuses on letting businesses detect whether they're talking to a bot at all. PACT assumes the business already knows it's talking to an agent and instead asks what that agent is entitled to do on its principal's behalf, and how the business can check that claim before acting on it.
Decagon says it's joining the Personal Agent Protocol working group that grew out of Meta's Muse team, which suggests these efforts are converging rather than competing outright — though it's early, and nothing here is yet a ratified standard with multi-vendor enforcement. The open question is whether scopes defined independently by thousands of individual businesses end up interoperable in practice, or whether personal agents end up needing a translation layer just to make sense of everyone's bespoke permission vocabulary. That's the part worth watching as PACT moves from a GitHub release to actual production traffic.
For now, what exists is a real, inspectable specification tying consent scopes to an existing discovery mechanism (the A2A Agent Card) and an existing auth flow (OAuth 2.0) — a concrete, if early, answer to the credential-scoping half of the personal-agent trust problem.