Current thinking

Imagine giving an agent access to a wallet to make a payment. I described a version of this on The Architects in January. A person could set a spending limit and specify what they want the agent to optimise for, such as cost or time. The agent could then look for liquidity or a route to complete the payment. The person would need to see what happened and have a way to resolve a dispute.

The same problem appears on the other side of the transaction. At ETHIndia, I used a driving licence to illustrate it. A counterparty needs a way to tell whether the issuer has revoked a signed credential. A directory can help verify changes to public attestations. Private credentials raise a separate design question about how much information the counterparty needs to see.

An open commerce network gives us a concrete way to think about settlement. A buyer’s agent might discover a seller across platforms and, if the buyer is less confident in that seller, create a smart-contract escrow. I offered that as a possible use of programmable settlement. It shows how a single purchase could involve discovery, verification, consent and payment across several systems.

A financial record can be digital and still require a series of calls, forms and agreements to use. On The Architects I described two properties that matter for tokenisation. Verifiability means another party can check a claim about an asset or its ownership. Transactability means the parties can use that asset in a transaction, subject to the relevant rules. The two properties together can make it easier to work with assets across institutions.

Take a security that someone wants to pledge as collateral. A digital statement can help the borrower show they hold the security. A system that can verify the holding and lock the asset under agreed terms could help another provider offer the loan. This still requires legal recognition of the rights involved, as well as controls for eligibility and flows across jurisdictions.

That difference matters for AI agents. An agent could help someone compare options today. Executing the chosen option requires each necessary step to be available in a form the agent can use within its mandate. The Finternet architecture work with Nandan Nilekani and Pramod Varma provides a way to think about those assets, institutions and connections.

At ETHIndia I described putting coordination logic before organisational logic and used payments to illustrate the idea. One starts with a clearing house, settlement agency and teams responsible for the work. Another uses shared rules, incentives and cryptography to let participants coordinate directly. Organisations remain important in either case. I’m interested in how much more the connections between them could do as agents become available to each participant.

The hourglass model that Abhishek Sankritik and I explored in a World Bank background paper gives this a practical starting point. A few shared standards at the narrow part let different technologies connect underneath and a wide variety of services grow above. The design choice is which rules everyone needs to share. Keeping that layer small can give participants room to develop different applications and change the technology over time. In the World Bank seminar, I also argued that AI could ease the work of adapting interfaces, while legal agreements and incentives still shape whether institutions will connect.

At ETHIndia I asked what such coordination might make possible in areas like drug discovery or agricultural supply chains. Those are questions for experimentation. Each brings different actors, incentives and approvals. An agent might help discover the relevant parties or assemble a sequence of commitments. The people and institutions responsible for each decision would still need a way to agree, verify and enforce what happens.

When an agent can move value, some of the rules governing that action need to be verifiable as the transaction happens. At the launch of the C:\>DIR Global Agentic Regulator Hackathon, I described why this matters for trust. A payment can have a purpose, a permitted recipient and conditions on how it can be redeemed. The people issuing and using it need to know how those conditions are applied.

Vouch gave the hackathon a sandbox for exploring this. Teams could work with rules for issuance, identity-linked access, redemption and audit visibility. A voucher offers a tractable example because it makes the question specific: who may receive it, where may it be used, and what record should be available when something goes wrong? The hackathon demos were prototypes for examining these questions.

I made a related argument in an earlier essay on technology-aware regulation. Regulators need to see how technical design affects risk and have a place to test how rules work before relying on them. A programmable rule can help with routine enforcement. The institution setting it still needs to decide how to handle exceptions, explain decisions and provide recourse.

My earlier work on data governance began with a fairly ordinary problem. A person generates information across many services, and that information can be useful to them elsewhere. Sharing it safely takes more than a convenient interface. In our BIS paper we proposed a system in which consent can be specific, revoked and audited. The person needs a meaningful say over where their data goes and what it can be used for.

In the World Bank seminar in 2025, I explained how work on consented data sharing informed my thinking about the Finternet. A common architecture could help individuals and businesses use identity, credentials and assets across services. The technical design has to work alongside legal recognition and a reason for participants to adopt it.

At ETHIndia I asked a more personal question. As AI takes on some kinds of cognitive work, how do we stay attentive to curiosity, feeling and the things we want to create? That question sits alongside the technical ones. I want to understand how people could delegate more of the work of coordination while retaining the ability to make choices, change their minds and decide what is worth doing.