Agent access to bank accounts becomes a protocol fight

The three ideas face very different competition. The Agent Traffic Census is already taken. HUMAN Security, Fingerprint, Arkose Labs, cside and Cloudflare (through Web Bot Auth with Visa and Mastercard) already classify agent traffic by operator, trust level and intent. They sell to the same merchant and fraud teams. A new entrant would need a very narrow angle, such as bank login pages with protocol-level evidence, and the edge and bot vendors could copy that quickly. The Poppy Gateway is the most open of the three, but the window is short. Poppy only reached Draft 0.1 in October 2026, with no reference implementation, no official SDK and payments not yet in scope. I found no hosted Poppy implementation for banks. The risks are concrete. Sierra could host it itself. Big banks in the design group, such as Bank of America, Wells Fargo and Chime, will build their own. Plaid is a partner and already sits between agents and US bank accounts. Agent-identity vendors like Gravitee, cidaas and Entersekt, plus open banking API vendors like Axway, already sell consent, scope and audit tooling to banks and can add Poppy support. The way in would be to support several protocols at once (Poppy, FDX and 1033 APIs, PSD2 consent, and Visa TAP and Mastercard Agent Pay signatures) for smaller US banks. Those banks also have to meet CFPB 1033 deadlines and have no in-house platform team. This only works if the product ships before the incumbents add Poppy. The Agent Liability Registry has a real, documented gap. Worldpay and other industry commentary say dispute liability between issuer, merchant and agent operator is still largely unresolved, and dispute tooling is mostly missing. Visa appears to push liability onto the cardholder, while Amex covers verified agents. Trulioo's Agent Directory, Skyfire KYA and AP2 mandates already cover agent identity and signed intent. What nobody seems to offer is a neutral record that ties agent identity, the mandate and what actually happened, and then feeds into chargeback reason codes. The risk is that card networks write this into their own rules and keep the evidence themselves. A startup would need to sit alongside chargeback and dispute vendors, or work with issuers on non-card rails like pay-by-bank, where the networks don't control the evidence.

• Personal Agent Protocol (Poppy) by Sierra and Meta

  The protocol this opportunity is built on. It is an Apache 2.0 draft built on OAuth and published in early October 2026. It has about 35 design partners, including Visa, Mastercard, PayPal, Plaid, Bank of America, Wells Fargo and Chime. A reference implementation was promised but not yet released, and payments are listed as a future extension. Sierra is well placed to ship its own hosted version.

• BBVA joins Poppy

  BBVA is the first European bank in Poppy. It has given no launch date. Separately, it completed a Visa Intelligent Commerce agent payment in July 2026. This confirms that the 'European banks following BBVA' segment is real but still exploratory.

• Gravitee (agent identity and API gateway for banks)

  An API management vendor pitching agent registration as OAuth clients, consent-scoped PSD2 and open banking access, and per-agent audit for banks. It is the closest existing match to a hosted gateway, though it is not Poppy-specific.

• cidaas AI agent identity

  A European CIAM vendor. It offers revocable per-resource consent, human sign-off for sensitive actions and tamper-resistant audit trails for agent actions.

• Entersekt Trusted Agent

  A bank authentication vendor that frames agent authority as traceable to an approved principal and enforced at runtime. It already sells to banks.

• Axway Amplify Open Banking (e.g. Regions Financial)

  US regional banks such as Regions are building FDX and CFPB 1033 APIs on vendors like Axway. A third-party rating said Regions had no agent consent or bot-identity features yet. These incumbent API vendors are the natural channel competitors for a regional-bank gateway.

• HUMAN Security AgenticTrust / Sightline

  Identifies agents by name, provider and type. It assigns trust levels, verifies agents cryptographically and lets customers set allow, read or checkout policies. This is almost exactly the Agent Traffic Census, and it is already sold to commerce teams.

• Fingerprint Authorized AI Agent Detection

  Launched in February 2026. It identifies signed agent traffic from OpenAI, AWS AgentCore and others for fraud and commerce use cases.

• Cloudflare Web Bot Auth with Visa TAP and Mastercard Agent Pay

  Signed agent identity at the edge. Its tags say whether an agent is browsing or paying, and it can link an agent to a consumer. Adyen, Shopify, Stripe and Amex are involved. Cloudflare can classify agents on login and checkout pages by default.

• Arkose Labs agent classification

  A fraud vendor that sorts agents into self-disclosing, legitimate non-disclosing and malicious groups. It markets this directly to merchant fraud teams.

• cside AI agent detection

  A smaller vendor with a five-class agent taxonomy and default allow or rate-limit actions.

• Trulioo KYA and Agent Directory

  Signed agent credentials plus a public Agent Directory that shows each agent's trust score, consent scope and status. It partners with Worldpay. This is the nearest thing to an agent registry, though it covers identity and consent rather than dispute allocation.

• Skyfire KYA

  Coined 'Know Your Agent'. It verifies the principals behind agents and issues tokens tied to their verification level for agent payments.

• Google AP2 (supported by PayPal)

  Cryptographically signed intent and cart mandates. This protocol-level evidence is what a liability registry would need to store and adjudicate.

Checked 2026-10-15