Developers post live apps, integrations, or demos.

Ecosystem Showcase & Launches

The UK Is Not Consulting on AI Agents. It Is Testing Three Architectures in a Sandbox.

A consultation asks what you think. A test rig finds out what breaks.



Most of what gets written about agent regulation is about the first kind of document. Someone will publish a paper, there will be a comment period, and in some number of years a rule will arrive. That is the shape everyone expects, and it is why most people in agent payments are building as though the regulatory question is still open.

In the United Kingdom it is not open in that way. It is being run as an engineering problem, with a budget, named researchers, and a sandbox.

What the 

Financial Conduct Authority's Smart Data Accelerator

 actually did was commission architecture research. The partners are Raidiam, which already serves as technical delivery partner for the FCA's smart data work, and the Saïd Business School at the University of Oxford. The work is led independently by Professor Pinar Ozcan, director of the Oxford Future of Finance and Technology Initiative, with researcher Kyeyoung Shin. John Heaton-Armstrong leads for Raidiam.

The output is not an opinion. It is a set of architectural models, translated into technically coherent blueprints, and tested in the FCA Innovation Department's Digital Sandbox against synthetic datasets.

Read that last sentence again. A financial regulator is not asking the industry what the architecture should be. It is building candidate architectures and trying to break them in its own environment.




The four dimensions, and the one that should stop you 👀



The research covers four dimensions. These are the parties' own words, from Raidiam's published description of the programme:

Infrastructure and standardisation, covering ecosystem rails, APIs, interoperability and resilience.

Technology, covering agentic AI, blockchain or distributed ledger technology, and quantum-readiness.

Trust and consent frameworks, covering consumer control, ethical data use and data assurance.

Governance and liability frameworks.



The second one is the sentence worth sitting with.



Agentic AI is not in a separate futures paper. It is not in an AI strategy document, or a horizon-scanning annex, or a speech. It is one of four technology considerations inside an infrastructure brief, listed between distributed ledgers and quantum-readiness.

That placement is a judgement, and it is a serious one. Quantum-readiness is in that list because cryptography has a replacement cycle measured in years and you cannot start it late. Putting agentic AI beside it says the same thing about agents: this is a property the plumbing has to carry, not a product category to watch.

It is the least glamorous place a regulator could have filed AI agents. It is also the most consequential.




Three architectures, and why testing all three is the point



The research develops multiple architectural models: centrally coordinated, federated interoperability, and hybrid.

It would have been cheaper to pick one. Most infrastructure programmes do, usually the centrally coordinated one, because a central operator is easy to govern, easy to audit, and easy to point at when something fails.

Building all three and testing them is a different posture. It treats the choice as unknown, which it is, and it means the question being answered is narrower and better than "what should the rules be". The real question underneath is which functions need common coordination, and which are better left distributed.

That question has an engineering answer rather than a political one, and it is the question anybody building agent infrastructure is already answering by their design choices, whether or not they noticed.

Here is the part that should interest anyone building in the open: the federated case has already been specified, in public, by people who had no choice.

 Nobody writing agent payment standards had a central coordinator available. There was no registry to join, no scheme to be accredited by, no hub to connect to

. So they had to describe how two parties who have never met establish enough trust to transact, with no middle.

ERC-8004 describes how an agent carries identity and a record. ERC-8183 describes what happens to the money when a paid job does not complete. x402 moves value between software at amounts a card rail will not carry. x401 describes who authorized the agent.

We have written the full map of those four before and will not repeat it here. The relevant point today is narrower. A regulator is now funding the work of comparing coordinated, federated and hybrid models for a regulated ecosystem. The federated column of that comparison is not a blank sheet. It has running code and public specifications, written by contributors from the Ethereum Foundation, 

, Coinbase, Google and the card networks.




About the date, because yesterday's date was a prediction



We wrote yesterday about a prediction, published by a card network, that at least three G20 regulators would issue formal guidance on AI agents by 2030. A prediction, written by an independent author, and we were careful to say so.

This is a different kind of object, and the difference matters.

The FCA published its open finance roadmap on 14 April 2026, setting a path to extend open banking principles across mortgages, insurance, pensions, investments and SME lending by 2030. In it, the FCA commits to working with HM Treasury on options for a regulatory framework for open finance by the end of 2027.

That is not a forecast about what a regulator might do. It is a published working date from the regulator itself, and it is roughly three years earlier than the prediction.

Be precise about what it covers, because the temptation to overclaim here is strong and we are not going to. The end of 2027 date is about open finance. It is not a date for a rule about AI agents, and nobody has announced one. Agentic AI sits inside the infrastructure research that feeds the open finance work. The honest statement is narrow and still useful: the architecture that will carry agent access to regulated financial data is being specified and tested now, against a framework timetable that runs to 2027, in the one jurisdiction that has already done this once with open banking.




What is not confirmed here, said plainly 🤔



A research note dated 29 September 2026, titled Powering open finance infrastructure, has been circulating. It describes the three model families in detail, names common foundations including OAuth 2.0, OpenID Connect, FAPI 2.0 and Strong Customer Authentication, works through international comparisons from Brazil, Australia, the European Union and the United Arab Emirates, and raises OpenID Federation as a possible future pattern for carrying verified information about an AI-enabled service's authority across trust domains.

We read that note through a relay rather than from the regulator's own site. We could not reopen it on fca.org.uk, where the research notes path returns a 404 and no search surfaces it. It is two days old, which is the most likely explanation, and we are telling you rather than implying access nobody gave us.

What we did confirm independently, and what the argument above rests on, is the commission itself, the named parties, the four dimensions, the three model families, the Digital Sandbox testing with synthetic data, and the roadmap dates. Those come from Raidiam's own published description of the programme, from trade coverage of the commission, and from the FCA's own roadmap.



If the note is what it appears to be, the detail is richer than the spine. The spine is enough.

An architecture is a diagram until value moves across it, and value does not move usefully without a unit of account a finance function can hold.

Arc's mainnet opened on 16 September 2026

. It is Circle's Layer 1, engineered for institutional finance, and its most important design choice for agent payments is the dullest one available: 
USDC is the native gas token. An agent settling a payment never has to acquire a second, volatile asset to pay the network fee. One asset in the account, one line in the books.

At agent scale that is structural rather than convenient. Software making thousands of small payments a day cannot run a treasury policy for a token it never wanted.

Arc launched with more than a hundred enterprise integrations and a validator consortium that reads like a list of institutions that do not join things lightly: BlackRock, Visa, Mastercard, DTCC, Standard Chartered, Intercontinental Exchange, MoneyGram, SBI Group, Sumitomo, Galaxy and Worldpay. More than 649 million dollars in USDC was bridged onto Arc in its opening week.

Around it, the rest is assembled. The Circle Payments Network reached 14.7 billion dollars in annualized volume with 175 enrolled financial institutions at the end of the second quarter. Circle Mint Europe issues both USDC and EURC from Paris under an ACPR licence, which is the relevant fact for anyone reading this from inside a European perimeter. Circle's national trust bank charter received final approval from the United States Office of the Comptroller of the Currency on 10 July 2026. And on the agent side specifically, 99.3 percent of observed x402 transaction value settles in USDC.



When software pays software today, it overwhelmingly pays in Circle's dollar.



We have been on Arc since the first day of testnet, 28 October 2025, and we have never deployed to another chain.

Meridian is one connected loop rather than a single product. Meridian Pay accepts payment in regulated digital dollars and euros. YieldVault puts idle balances into tokenized Treasury bills, with no fixed or guaranteed rate offered or implied. Payment Streaming pays by the second rather than at month end. And Firmata Protocol answers who is on the other side before the money moves.

Firmata composes ERC-8004 for agent identity and reputation with ERC-8183 for conditional escrow, settled over x402. It performs no anti-money-laundering or know-your-customer checks and replaces nobody's obligations. It publishes and verifies attestations about a counterparty, and it holds funds against a condition. That is the whole claim and we will not inflate it.

Our stack runs on Arc testnet. Twenty contracts, each verified on Arcscan from the second of deploy. An independent security audit gates the move to mainnet and we have not made it.





This is commissioned research, not regulation. The FCA paid for it and Oxford ran it independently, which the FCA itself is careful to state. No rule has been proposed about AI agents in financial services in the United Kingdom, and nothing above should be read as saying one has.

The perimeter is British. Open finance in the United Kingdom is not open finance in the European Union, where PSD2 and the Financial Data Access framework apply, and it is not the United States, which has taken a different route entirely. A model tested in a London sandbox is evidence about London.

Research can be shelved. Plenty of well-run public research produces a document and no consequence. The reason to pay attention here is the sandbox rather than the paper: a blueprint that has been run against synthetic data in a regulator's own environment is harder to ignore than a recommendation, but harder to ignore is not the same as adopted.

And we are not neutral. We build agent trust infrastructure on open standards and in a federated model. Research that takes the federated case seriously is research that flatters our architecture, and you should discount our enthusiasm accordingly. That is why the source status above is spelled out rather than summarized.






The usual advice about regulation is to wait for clarity. That advice assumes the clarity arrives as a document you can read and then implement.

This one is arriving as a comparison. Three architectures, four dimensions, tested side by side, in an environment built for breaking things. Whatever survives that becomes the reference the next document is written against, and that happens well before anybody publishes a rule.

The federated column of that comparison is the one that is already written, already running, and already settling value in public. Not because its authors were ahead of a regulator, but because they never had a central coordinator to lean on and had to specify the hard case first.

A consultation asks what you think. A test rig finds out what breaks.

The second one is further along than most people building in this space have noticed.

Built on Arc. Powered by Circle.



Not investment advice. Nothing here forecasts the decisions of any regulator.

🧪 The UK Is Not Consulting on AI Agents. It Is Testing Three Architectures in a Sandbox.

Home

Forum

Events

Content

Courses

Collections

People

Contributors

Albums

Groups

Chat

Channels

Not being used

Help

Join the global network of builders developing the Economic OS for the internet. We’re reimagining how value moves, creating the tools that make the economy work for everyone.

Arc House

What is your job title?

Join a global network of builders developing the Economic OS for the internet. We’re reimagining how value moves, creating the tools that make that economy work for everyone. 

Arc House | Build the economic OS for the internet with us

Copyright

Terms of Service

Privacy

Code of Conduct

Your Privacy Choices

It looks like there is no account associated with the email you provided. The event is for members only. Please complete the application to join our community first.

Thank you for your interest in joining us

Thanks for your interest in joining the event. However, your account is still pending. Once it's approved, you'll be able to register for the event. Please contact us via community@arc.io if you need support.

Pending Approval

Login to the The Arc Community Hub community to connect with others, attend community events, and more!

Login or Sign Up for even more

To unlock all parts of the community complete your profile using verifiable and real information.

Hey, complete your profile

It looks like there is no account associated with the email you provided. Please complete the application to join our community.

Thank you for your interest in joining Arc House.

Thanks for your interest in joining Arc House. We will review your request shortly and email you once it's approved. Meanwhile, please feel free to contact us via community@arc.io for any questions.

Accounts are unapproved when profile information cannot be verified or when activity patterns indicate spam, inorganic participation, or behavior that undermines trust in the program.

To strengthen your profile for future consideration, sign in with a verified LinkedIn account, use real and current profile information, and participate constructively in other community spaces.

Commercial customers should raise concerns through their support channel.

https://www.arc.network/community-guidelines