Your customers’ AI agents
are already logged in
AstraSync shows you which sessions an agent is driving, gives good agents a declared way in under a mandate the customer sets, and keeps a record a regulator can check.
~12 ms
median verdict, before the trade and outside the matching path
8 of 8
MAS SAFR requirements demonstrable, by our own assessment
0
payment credentials ever held by the agent
1
named, identity-verified human behind every agent
Agents are already inside authenticated sessions
Robinhood opened a sanctioned door for agents and counted 150,000 agentic accounts in four months. Everywhere else the same demand arrives through an ordinary API key or a browser session, and nobody is counting.
Banks
The agent walks in with the customer
Browser agents run inside the customer's own browser: same device, same IP, same cookies. The customer passes MFA, then hands over the mouse. Every fraud signal says it is the customer, because for the first ninety seconds it was. Nobody asks whether the customer is still the one driving.
Exchanges & brokers
The algo you approved became an agent
Venues certified deterministic scripts behind an API key or FIX session. Swapping in an LLM-driven agent that reads the news and reinterprets its instructions needs no new key, no new session and no notice. From the venue’s side, nothing changed.
Expect three questions after the first complaint reaches a regulator.
- 01Could you tell an agent was operating the account?
- 02What did the customer authorise the agent to do, and can you evidence the scope?
- 03What have you changed since you found out?
“Our terms prohibit it” and “our logs say it was the customer” are both true. Neither is a control.
Let good agents in, with a verified human behind each one
One checkpoint for every agent that reaches you: on a trading venue, in a banking app or at a merchant checkout.
Exchanges & trading venues
A checkpoint in front of the execution engine
This year we have worked with an institutional cross-asset exchange, regulated in multiple tier-1 jurisdictions, that wanted AI agents as clients from the outset. It never asked how to keep agents out. It asked how to let good agents in, with a named and verified human accountable for each one.
- Every agent resolves to an owner who has passed identity and AML checks
- Every order is tested against a machine-readable mandate: purpose, duration, limits and scope
- Three outcomes: execute, deny with a recorded reason, or hold for a human with a timeout that defaults to block
- Action, scope and value are read from the live request, never from what the agent says about itself
- Median verdict around 12 ms, outside the matching path, with no change to the engine
Validated in the venue's integration environment and on a scoped path to production. Not yet carrying real-money flow.
Banks · Bring Your Own AI
Let customers bring their own agent, on your terms
Your customers will use agents whether you sanction them or not. Bring Your Own AI gives them a declared path in: the customer activates their agent from inside your app, against the identity you have already verified, and decides what it may do.
“Bring Your Own AI using a secure verification gateway.”
- No re-onboarding: the bank attests the customer it already knows, so activation is a tap in the app
- A bounded mandate per agent: which accounts, which actions, which limits
- Disputes, transfers and other sensitive actions are held for the customer to approve
- Revoke any agent from the banking app, effective on its next request
- A white-label activation widget that sits inside your existing app
Working end to end today. We will walk your team through it on a demo bank.
Card issuers & lenders · Agent payments
Your card in the agent's checkout, without handing the agent the card
When a cardholder asks an agent to buy something, the agent never sees the card. AstraSync issues a single-use authorisation from the cardholder’s own card that only the merchant can redeem, and your brand stays on every screen.
- The agent holds no card number, token or credential at any point
- Purchases above the cardholder's limit pause for their approval, right in the chat
- Your preferred merchants first, with cardholder pricing or interest-free terms shown at the point of sale
- Not a cardholder yet? Apply for the card inside the purchase
- Your brand on the activation, the cards and the receipts, powered by AstraSync
Available as a white-label issuer programme. AstraSync never charges or settles: your rails do.
See, declare, keep the evidence
The first step is cheap and needs no customer friction. Each step maps to a product you can switch on independently.
See
Classify what is already arriving: human, scripted or agent. No customer friction and no policy decision. It turns "none that we know of" into a number a risk committee can work from.
Visit IntelligenceOffer a declared path
An agent only identifies itself if doing so gets it something. Let customers register an agent against their verified identity and grant it a bounded mandate it cannot reason its way around.
Trusted Agent Gateway · Bring Your Own AIKeep the evidence
A record made at the time, which an outsider can check without trusting us: which agent, acting for whom, under what mandate, did what. That moves the conversation from "the logs say it was you" to "here is what you authorised".
Compliance ControlMapped to MAS SAFR, part by part
In July the Monetary Authority of Singapore published SAFR, Safeguards for Agentic Finance at Runtime. With the exchange we work with, we submitted a joint response mapping a working deployment to it.
SAFR’s four working parts
- Registry
- Every agent resolves to a person who has passed identity and AML checks
- Machine-readable controls
- Mandates across purpose, duration, limits, scope and sub-agent spawning
- Decision engine
- Allow, deny or escalate to a human, with no approval carried to the next action
- Audit log
- Tamper-evident and anchored on a public blockchain, so a supervisor can verify it
By our own assessment all eight SAFR requirements can be demonstrated, and we invited MAS to validate that independently. The response supports an expression of interest in MAS’s industry work group and is not an endorsement by MAS.
Existing rules already apply
FINRA
Its 2026 oversight report names the risk of an agent acting beyond the user’s intended scope and authority.
Federal Reserve
Governor Waller: the task shifts from identifying the customer to proving an agent holds the customer’s authority.
APRA
Identity and access management has not adjusted to non-human actors, and existing standards already apply.
Bank of England
Its frameworks were not built to contemplate autonomous agents.
Every agent's owner is AML-screened
Trust in an agent starts with the human behind it. Owners pass anti-money-laundering screening before their agents carry a verified identity, and they stay screened.
Sanctions, PEP & adverse-media screening
Every verified owner is checked against global watchlists (OFAC, UN, EU, HM Treasury, DFAT) plus PEP databases and adverse media.
Continuous monitoring
Screening isn't one-and-done: owners stay under automatic AML monitoring, and any risk change propagates to their agents immediately.
Attestation-gated access
Platforms can require fresh AML attestations, like a PEP-clear check no older than 7 days, before an agent is allowed in.
Trust-score impact
Completed KYC/AML feeds the identity block of the trust score, so compliant owners give their agents a measurable head start.
Find out how many agents are already in your sessions
Start with an Agent Traffic Audit in observe-only mode. No customer friction, no policy change, and a number your risk committee can act on before the first complaint does it for you.