Subscribe to get high-signal insights on how modern fintech is built.

opinion

What AFSP v0.1 Won't Let Your Agent Do

The first agent-to-bank protocol requires three minutes of cryptographic handshaking to open a savings account.

By Alex Kugell ·

The first formal standard for how AI agents talk to banks is 57 pages long. It covers one use case: opening a savings account.

AFSP, the Agentic Financial Services Protocol, went to public review on September 28. The spec is architecturally serious.

FIDO2 biometric consent on enrolled devices. Hardware TEE-signed authorization tokens. ECDSA-P256 cryptography. A five-signal attestation package the agent must present before the bank sees a single API call.

All of it for the lowest-risk operation in consumer banking.

I think the protocol gets the trust architecture right. I also think starting with account opening tells you exactly how far the industry is from letting agents touch money.

Five signals for a savings account

Before an agent can open an account on your behalf, it has to assemble what AFSP calls a Pre-Credential Agent Attestation. Five signals, bundled into a cryptographically signed package, verified before any API access.

S1 is agent provenance. A certified, registered, unmodified build from an accountable operator. The operator registers the agent in an AFSP certification registry, except that registry, AFSP-07, is still on the roadmap.

S2 is biometric consent. You authorize the session on your enrolled device using FIDO2. Your phone's secure enclave signs a token that locks down what the agent can do: which action classes, which product categories, what time window. The agent cannot expand its own scope.

S3 is a federated referral. Optional. A bank you already have a relationship with vouches for your identity, sharing only the data you have explicitly permissioned under Section 1033.

S4 is financial identity. Behavioral signals consistent with an established financial profile. Transaction history that matches who you claim to be.

S5 is session integrity. Conditional. Is a human actually operating this session, and does the behavior fall within normal parameters?

The whole handshake takes about two minutes and fifty seconds. Then the agent opens a savings account. The session ends.

Pre-Credential Agent Attestation
S1Agent ProvenanceCertified, registered, unmodified buildrequired
S2Biometric ConsentHuman authorized on enrolled device via FIDO2required
S3Federated ReferralPrior KYC from a trusted institutionoptional
S4Financial IdentityBehavioral signals match established profilerequired
S5Session IntegrityHuman-operated, normal behavioral parametersconditional
PCAA Package
ECDSA-P256 signed · nonce-verified · validity window enforced
Bank API · Account Opening
Savings account opened · ~2 min 50 sec

What the roadmap reveals

The weight of PCAA makes more sense when you see what AFSP plans to do next. Six more components beyond v0.1. Three of them point directly at where the hard problems live.

AFSP-09 is Pre-Execution Revalidation. Before each action in a session, re-verify that the authorization still holds. This matters when the agent's next action is a $50,000 wire transfer. It does not matter for a savings account application.

AFSP-10 is the Behavioral Envelope Engine. Real-time monitoring of whether the agent stays within the boundaries the human defined. You authorized "open a high-yield savings account." The agent starts browsing brokerage products. Something needs to intervene.

AFSP-12 is Post-Opening Session Continuity. What happens after the account exists? Can the agent fund it? Move money in? Set up recurring deposits? v0.1 does not say.

These three components mark the line where identity stops being the problem and authorization begins. Proving who you are uses mature tools: FIDO2, biometrics, device attestation, cryptographic signing. That infrastructure works. Defining what an agent can do with your money, monitoring it in real time, deciding who absorbs the loss when it exceeds its scope: none of that has shipped.

AFSP v0.1 builds the identity layer and stops at the edge of the authorization layer.

AFSP coverage: shipped vs. roadmap
Identity
AFSP-01 – 06
Pre-Credential Agent Attestation
Five signals, PCAA assembly, account opening
v0.1
Authorization
AFSP-07
Agent Certification Registry
Who certifies agent builds and maintains the registry
roadmap
AFSP-09
Pre-Execution Revalidation
Re-verify authorization before each action in a session
roadmap
AFSP-10
Behavioral Envelope Engine
Real-time monitoring of agent scope boundaries
roadmap
AFSP-12
Post-Opening Session Continuity
What the agent can do after the account exists
roadmap

The Robinhood contrast

While AFSP assembled its attestation framework and opened a public comment period, Robinhood shipped.

At Hood Summit 2026, Robinhood announced agent integration through MCP server endpoints. Since May, 150,000 customers have connected AI agents to their brokerage accounts. Those agents process roughly 30 million tool interactions per day. Free until year-end, with a subscription or per-agent fee expected in January.

Robinhood built MCP endpoints, let agents call them, and wrapped rate limits and existing risk controls around the edges. The entire integration shipped without a protocol spec, an attestation framework, or a governing foundation.

AFSP says: standardize the trust layer before agents access financial systems. Robinhood says: ship the integration, instrument everything, learn from production traffic.

Neither is obviously wrong. AFSP's caution makes sense when the failure mode is an agent draining a checking account. Robinhood's speed makes sense when the agent operates within risk controls Robinhood already enforces for human traders.

But the gap between the two approaches is measured in years. AFSP's governing foundation will not form until early 2027. The public comment period on v0.1 runs through November 16. By the time AFSP-09 through AFSP-12 are ratified, Robinhood will have processed billions of agent interactions and built its risk model from observed behavior, not protocol specification.

What AFSP cannot solve alone

AFSP is authored by Primitive, a startup serving as interim protocol administrator. Visa, Mastercard, and Ant International are building their own agent trust framework. ACP, AP2, and x402 are competing on agent payment authorization. The compliance stack is assembling its own layers. Policy engines are shipping independently.

Multiple groups building different slices of the same puzzle. No single authority deciding how they connect.

AFSP's slice is specific and well-built. The four-party trust model puts control where it belongs, with the consumer, who retains sovereign authority and approves every step on their own device. Tamper-evident audit trails are retained for seven years. Examination-ready from day one.

But account opening is the operation that needs this infrastructure least. Who is liable when an agent buys the wrong thing? What replaces strong customer authentication when the payer is software? How does dispute resolution work when there is no chargeback code for "my agent hallucinated"?

I have been tracking these questions for months. None of them have a protocol answer yet.

Fifty-seven pages of protocol for a savings account is a strong foundation. Nobody has written the spec for what happens when the agent needs to move $50,000. That version of AFSP is the one worth watching.

Sources

Alex Kugell
Alex KugellAI/ML Engineer, Fintech Infrastructure

Writes Ledger Drift. Engineer building AI systems and financial infrastructure. Open source contributor to x402, ACK, and MPP – agentic payment protocols.

Frequently Asked Questions

What can AI agents actually do through AFSP today?
Open a deposit account. Nothing else. Payments, transfers, credit applications, and trading are outside the v0.1 spec. The protocol requires five cryptographic attestation signals, including FIDO2 biometric consent and hardware-signed tokens, just for that single operation. Pre-execution revalidation and behavioral monitoring are on the roadmap but haven't shipped.
Why does AFSP need five attestation signals to open a savings account?
The protocol bundles biometric consent, hardware TEE signatures, platform identity, consent scope, and behavioral monitoring into a Pre-Credential Agent Attestation. Each signal addresses a different attack vector: proving the human exists, that the device is genuine, that the agent is registered, that the scope was authorized, and that runtime behavior hasn't deviated. Starting with the most conservative trust model lets the spec relax controls later rather than patch them after an incident.
How are companies like Robinhood handling agent access without AFSP?
Robinhood shipped MCP endpoints directly, onboarding 150,000 agent accounts with 30 million daily tool interactions since May 2026. The approach treats agent access as an API integration problem rather than a protocol-level trust problem. The tradeoff is speed versus formalism: Robinhood has agents in production while AFSP is still in public comment, but Robinhood's approach depends on its own platform controls rather than an industry standard.
When will AFSP support payments and transfers?
No timeline has been published. The v0.1 spec covers account opening, and the AFSP Foundation that will govern the standard isn't targeted to form until early 2027. Public comments on the current version are open through November 16, 2026. Starting with the lowest-risk banking operation and building toward higher-risk ones means payments could be years away at this pace.
What does AFSP's scope tell you about the state of agent banking?
A 57-page protocol spec for account opening, the lowest-risk operation in consumer banking, signals that the industry has no established trust model for agents touching money. Payments require real-time authorization, dispute handling, and liability assignment that are absent from the current spec. The gap between AFSP's careful approach and Robinhood's production deployment shows an industry split on whether to build the trust model first or ship and patch.

Built by Trio, a fintech-native engineering partner helping teams build the next generation of financial technology and infrastructure.

Subscribe to Ledger Drift for high-signal insights into how modern fintech is built, from systems to code to teams.

Keep reading

anatomyFragment Went from Ledger API to Financial Control PlaneFragment sold fintech engineers a managed ledger, and two years later Ramp's coding agents use it as their financial control plane.
analysisMost Stablecoin Payments Never Cross a BorderStablecoin payments hit $482 billion in nine months, and 61% of the volume that can be traced to a geography stayed inside the same country.
opinionAfrican Fintech: The ChoiceAfrica is building financial infrastructure from scratch, and the early returns look like payday lending with better UX.
View more ›