Agentic Economist

@sxtvik envisioning agentic commerce for consumers initially is so silly - shopping is fun.

@JillCastilla It seems like this should be solvable with SRF? I don't see why you couldn't extend SRF to be 24/7, allow for collateral outside of treasuries, and open it up to all depository institutions.

The payments system will only be as accessible as the discount window. We should be able to fund exchange money 24/7 with the ability to draw upon overnight funds that has collateral pledged. Otherwise, liquidity risk elevates with the speed of payments.

you ever attempt to negotiate payment terms? its a pain in the ass, takes a ton of time. payment terms sit in this land of opportunity where they are phenomenally flexible (vendors/merchants will happily agree to net-60 and net-90 if you're large enough and trustworthy enough). and there are many levers within a start MSA outside of the payment window that can all be negotiated.

@wealthofagents @CapuzR @KenWattana What do business want to get out from agentic negotiations? Speed? Better results? More supply? Would love to learn

Agentic payments gets people up in arms is it seems ephemeral and too good to be true Where is it being used in production? What are the use cases? Why haven’t I used it? Who are the agents paying? It’s a category that’s just a few months old. Good things take time. AI is rewriting how organizations do business and change happens faster than you think. And don’t discount the dozens of teams working behind the scenes to make it a reality!

x402 is a weird middle ground... fundamentally impossible to differentiate between a real consumer visiting your site vs an agent pretending to be a human. It relies on this assumption that agents will self-identify as such. if you believe all web traffic to your site will be agents from OpenAI/Anthropic (who at least seem to be honest with their User Agent) then its great and you should use it. if you believe all web traffic is real humans then just don't use it. but if you think you're still gonna get a mix of agentic and human traffic you risk ruining the experience for all humans involved.

@naturalpay very excited to see what natural comes up with

Payments made Intelligent. Natural takes care of complexity at the infrastructure layer. Verification, compliance, ledgering, integrations, and reconciliation happen beneath the surface, allowing builders to focus on what their agents are meant to do, rather than how money moves. Complexity disappears. Attention shifts back to innovation. Build intelligently with Natural: natural.co/blog/intellige…

i agree - limiting the conceptualization of agentic payments to card transactions is short-sighted. the ability for an agentic to negotiate net-60 vs net-30 on your behalf will have an outsized (positive) impact on SMBs. imo there will be a number distinct domains: - negotiations/contract terms - card payments (currently most popular) - invoicing/trade credit - ACH/bank transfer (highest risk imo) - procurement/p0 workflows - receivables/collections - compliance/reporting (maybe bucket audit trails in here as well)

@KenWattana @wealthofagents Real corporate volume runs on net-30 invoices, credit lines, and procurement workflows. How are you imagining corporate agentic spending? Tied to negotiations? policies based on B2B contracts? No x402 or cards heavily needed on that level probably..

@masonnystrom seems like an incredibly challenging problem given the unknown number of LLM calls, tool calls, and page visits required to execute a transaction - purchasing paper towels vs planning a trip to europe differ extremely in the amount of overhead

Agents paying per query / usage is interesting, but I think part of the problem of agent payment adoption is that no consumer thinks that way. Consumers don't want to have a variable bill. Consumers want to have full confidence that their ChatGPT costs $20/month. This leads to two types of Agentic payment models: Enterprise model - where usage and query based is fine and potentially preferred Consumer model – where a subscription are as part of the transaction fee is better (e.g. agent fee of $1 on flight booking). Agentic payment volume today is nascent. But it will obviously grow. Agentic payment tools building for consumer need to be opinionated about the payment experience for the end user and make it simple, seamless, and convenient.

@jaredhanson @bekacru @KeycardLabs Keycard is definitely one I've been keeping an eye on, I think the audit trail ability y'all are building will become critical.

@bekacru OAuth wasn’t designed for this, true. But there is a bunch of OAuth extensions that are. @KeycardLabs we are implementing all of this (and more) and making it available as a service!

I think we have a problem here because of how people think of what an agent is An agent isn’t the application (which oauth is designed for). An agent isn’t “cursor” or “claude code”, the agent is the specific actor within that runtime. Two separate chats in the your cursor are not the same agent. They have different contexts, different intents and should have different permissions, identities, and lifecycles. And unfortunately oauth was never designed for this An agent that is only supposed to read my email must not have permission to delete my email, even if I want another agent to be able to do that within a given time frame

@bekacru Yeah but this is just a matter of having appropriate fine-grained ACLs, not a critique of Oauth.

The real question around agentic payments is one of risk and liability - not the actual ability to transact. Will be interesting to see who succeeds in this space.

Gotta hand it to the Mastercard team, Verifiable Intent is such a good first foray into constructing a new trust paradigm for agentic payments. What I like: - The credential chain mechanics of the 3-layer SD-JWT architecture is really well designed. - Selective disclosure is genuinely useful <- I expect any other standards in this space to also go the selective disclosure route. There's literally no reason for each party to see everything. - Failing closed is the correct default behavior. But there are a couple shortcomings: - VI assumes the principal is a consumer, but neglects to conceptualize a business as the delegating entity. - No audit trail mechanism -> they sorta attempt to solve this with a vague hand-wavy `prompt_summary` field, but that is ripe for confusion. - There's no agent identity mechanism - cnf.jwk in L2 is just a key, its missing an attestation that the key belongs to a verified entity (whether consumer or business). In Mastercard's conceptualization, this is implicitly solved for by the chain L1->L2 i.e. "the entity that controls the L1-bound key chose to delegate to this agent". But its missing the concept of resolution; can a 3rd party look up what that agent is, do they know whether its an Claude instance or Gemini or what. The chain only proves authorization but not discoverability. It's a great piece of the puzzle but doesn't quite complete the picture and there are still some huge gaps around fine-grained ACLs, agent identity, delegation, etc.

@dansemperepico Yes but run it with this agent-safehouse.dev to limit its permissions then it just becomes `safehouse claude --dangerously-skip-permissions` and its great.

You guys all run Claude Code with claude --dangerously-skip-permissions right? Because otherwise how in the world can you sit there accepting every single permission when building something?