AI agents are getting payment credentials
Oct 20, 2026Tuesday, October 20, 2026Summary from 4 contributing podcasts.
- Card networks are creating limited-use credentials so AI agents can make purchases without exposing a customer’s card number.
- Bitcoin and stablecoin payment standards offer other routes, but responsibility for an agent’s mistakes remains unresolved.
- Merchants face a choice: block shopping agents or try to sell to them.
AI agents are moving from recommending purchases toward completing them. The emerging payment tools give agents a way to spend within limits, but they also raise questions about who controls a purchase, who handles a dispute and what happens when an agent gets it wrong.
On The a16z Show, VGS chief strategy officer Terry Angelos described card-network credentials that temporarily authorize an agent for a particular merchant and transaction limit, without handing over the customer’s primary card number. Angelos said VGS went from provisioning none three months earlier to a few million agentic tokens a week, as of his Oct. 15 account. The credentials aim to preserve familiar card protections, but Angelos said responsibility for an agent’s mistake, such as booking the wrong dates, is still unsettled.
Other payment routes are being connected to AI tools. Adam Curry said on Podcasting 2.0 that he had added multiple payment approaches to an agent connector for the Podcast Index. Sam Sethi pointed to x402, which enables payments in USDC, and L402, which uses Bitcoin’s Lightning Network. They are alternatives to card credentials, not the same system, and Curry expects the agent or its user to choose among payment methods.
Bitcoin advocate Dhruv Bansal, speaking on What Bitcoin Did, argues that agents may favor Bitcoin because banks or stablecoin providers can restrict access to funds. That is a prediction, not evidence that agents are already choosing Bitcoin. The contrast is between payment systems built around existing financial institutions and a permissionless network; the new card credentials do not settle that debate.
The commercial stakes are already visible. On The a16z Show, Alex Rampell and Angelos said Amazon blocked the Muse shopping assistant, arguing that agents can bypass the ads that help fund online retail. They said Walmart and Shopify merchants may instead welcome agents as another way to reach buyers.
In a later a16z discussion, Rampell and Jean-Denis Greze argued that automated shopping could also weaken ad-based business models and rewards tied to consumer habits. The spending tools are taking shape; the rules for access, accountability and revenue are still being worked out.
Source Intelligence
- Deep dive into what was said in the episodes
What Happens When Al Agents Can Spend Your Money? • Oct 19
- Jean-Denie argued that personal agents need incentives aligned with their users, since ad-funded assistants could steer recommendations toward advertisers. He also said agents should remove unwanted chores without eliminating enjoyable friction, such as browsing shops or discovering products through advertising.
- Jean-Denie said agents could weaken supply aggregators by removing the value of browsing and comparison, putting Amazon's advertising model at risk if shoppers buy through agents instead. He suggested retailers may need their own agents, while Meta could keep an agent free and monetize better targeting on Instagram.
- Alex Rampell pointed to CamelCamelCamel as evidence that shoppers already use agents' precursor tools to automate price watching; he said it has 20 million users in America. Jean-Denie added that automated coupon and credit-card benefit use could end breakage, changing businesses built on customers failing to claim rewards.
Also discussed on this episode: (13)
Startups (2)
- Jean-Denie Gres left Plaid to automate SMB tax returns, but found a weak venture case: the business made just under $1 million in year one and projected $2.5 million to $3 million in year two, while established tax firms could automate for existing customers.
- Jean-Denie said annual tax sales limited repeat use and word of mouth, while October 15 marked a narrow window to pivot before committing to another tax season. He and his co-founder changed direction because they wanted greater impact and a stronger future for their team.
Agents (4)
- Jean-Denie said Town emerged from customers asking for help with meeting preparation, follow-up, and routine emails. As coding models became more agentic, the team saw strong user demand and chose to build before fully sizing the market.
- Jean-Denie described Town as a way to surface work worth delegating to AI as token costs fall. Examples include preparing research notes for every meeting and checking restaurant suppliers' changing prices, tasks too tedious or costly to assign to a person regularly.
- Jean-Denie favors a horizontal product that learns from users, then offers role-specific automations rather than branding itself around one niche. Town interviews power users, builds reusable templates, and uses customer signals to recommend relevant workflows without inspecting private user data.
- Jean-Denie said Town can use agents as an internal company directory, finding colleagues connected to a prospect without waiting for employees to answer a Slack message. He sees this as faster coordination, but says companies need carefully designed privacy and security rules before trusting agents to mediate access.
Enterprise (4)
- Jean-Denie argued that workplace AI must feel as polished as a consumer app or risk becoming an enterprise-only tool people tolerate. He compared that danger to Good Technology's clunky iPhone email app and said Town prioritized the user experience before expanding enterprise administration features.
- Jean-Denie's adoption plan starts with email and calendars, which can reveal useful work without access to a company's core systems. Deeper integrations require trust, audit logs, privacy controls, and spending oversight; he expects smaller firms to adopt faster than large enterprises.
- Jean-Denie said AI adoption stalls because most businesses lack people who know how to tinker with it; users often treat chatbots as better search instead of building workflows. Town's opportunity is to show nontechnical users a useful automation directly, including unexpected uses across small businesses.
- Jean-Denie cited a surfboard rental business using Town to inspect returned equipment and decide whether to refund deposits, an example of AI use outside expected software markets. Alex Rampell added that a Harvard professor dismissed current AI after trying it years earlier, then changed her view after a new demonstration.
Models (3)
- Jean-Denie said model capability alone will not protect software companies as AI weakens the app layer. Town is betting that automated, reusable customization will scale better than sending human forward-deployed engineers to build bespoke systems for each large client.
- Jean-Denie pitches Town as a neutral layer that can route work to whichever model is best for a task, protecting customers from a supplier losing its edge or raising prices. He also says businesses should own their AI context, be able to switch quickly, and avoid model training on their data.
- Jean-Denie said models improve faster than many users update their expectations, making concrete demonstrations more persuasive than feature labels like computer use. He cited assistants completing browser tasks reliably now that they failed on most attempts months earlier.
When AI Agents Start Spending Your Money | Alex Rampell & Terry Angelos • Oct 15
- Terry Angelos says VGS secures credentials for about a third of e-commerce by GMV, serving major online merchants, neobanks, and frontier labs. Its vault lets merchants process stored cards without directly handling the underlying payment data.
- Card networks now issue merchant-domain-bound network tokens instead of exposing the card’s primary account number to each processor. Alex Rampell says VGS also helps merchants reduce PCI DSS obligations by keeping payment data out of their systems.
- Terry Angelos describes a recent shift from agents recommending purchases to completing transactions. Payment credentials give agents the ability to act, as in Angelos’s example of an agent finding and booking a piano teacher around his family’s schedule.
- Payment wallets are competing to connect with agents because agents need broad consumer coverage and merchant reach. Angelos cites Link at about 300 million consumer accounts and 1 million merchants, Shop Pay at about 1.7 million merchants, and PayPal at over 400 million accounts and roughly 30 million merchants.
- Wallets can store credentials for consumers to present at checkout, or handle the transaction themselves and remit funds to merchants; agentic commerce blends the two models. Alex Rampell says issuers have also paid tens of millions of dollars to make their cards top-of-wallet in services such as Uber and Airbnb.
- When Stripe Link could not check out at Walmart, Stripe bought the item with a one-time virtual card and charged Angelos separately. That worked, but Walmart and Angelos’s bank saw Stripe rather than each other, complicating follow-on subscriptions and disputes; card-network chargebacks provide a trust process that direct crypto payments lack.
- Card networks’ agentic tokens are designed to let wallets pass credentials through agents to merchants while preserving the consumer’s card relationship, dispute rights, and merchant follow-up. Angelos says VGS provisioned none three months earlier and now handles a few million tokens per week for agents, wallets, and merchants.
- Agent payment standards can verify that a card is valid, that an agent is registered with a card network and acting on a user’s instruction, and that a transaction fits its authorized context. Angelos says responsibility for agent mistakes, such as booking the wrong dates, remains unresolved.
- Merchants have different incentives toward agents: Alex Rampell says Amazon has reason to block them because advertising depends on shoppers’ attention, while Walmart may benefit from its scale and delivery network. Shopify’s small merchants may need agents as a sales channel; Amazon blocked Muse, and Rampell says Comcast blocked agent logins.
- Terry Angelos compares agent access to open banking: customers may want to delegate access, while platforms can restrict it to controlled APIs, limit the data shared, and charge for access. He points to the negotiated settlement between banks such as Chase and data intermediaries such as Plaid as a precedent.
- Angelos considers an agent-driven run on savings accounts overhyped because existing bill-pay integrations create friction, but sees unmonetized demand as underhyped. Agents could act on recommendations users forget to follow, such as buying safer cycling tires, while also finding and canceling unwanted subscriptions.
- Always-on agents could make it easier to exploit price thresholds and hidden terms, but could also undermine card rewards built around consumers spending across categories. Angelos argues that reward points and other breakage-based models may need new terms as agents optimize which card or offer to use.
- Angelos describes VGS’s progression from a credential vault, to managing credential updates and network tokens, to brokering credentials among multiple parties. That shift makes the vault multiplayer, much like shared password managers, and agentic commerce adds more participants to its network.
Also discussed on this episode: (2)
Payments (1)
- Large merchants use multiple processors to balance cost, approval rates, and regional coverage; Angelos says merchants above $1 billion in GMV average three to five. VGS maintains a neutral vault and helps keep credentials updated when cards change networks, expire, or are replaced.
VC (1)
- Alex Rampell and Terry Angelos founded TrialPay in 2006; Visa later acquired the company and invested in VGS while Angelos was at Visa. Rampell says he led an early VGS investment and now serves on its board alongside Angelos.
Podcasting 2.0
Adam Curry
Episode 275: "Off The Rails" • Oct 16
- Curry and Jones see app-to-agent connections as a practical way to combine personalized discovery with familiar playback. Curry says agents currently require explicit connectors and user approvals, while Jones says apps need a way to receive recommendations and instructions from a listener’s own agent.
- Curry says listeners’ agents could send boosts or donations directly to creators and payment splits, but app developers may still be excluded from the money flow. Jones sees no clear revenue model for apps under current advertising and contract arrangements, while Curry points to programmable payments such as L402 as a possible route.
Also discussed on this episode: (14)
Agents (3)
- Adam Curry says AI agents already personalize podcast discovery by using subscription history and other personal context, then recommending episodes across languages and topics. Curry argues RSS metadata gives agents a durable catalog to search and play through services such as Sonos.
- Dave Jones wants podcast apps to preserve a focused listening environment; separate inboxes and dedicated podcast apps help him keep tasks mentally distinct. He argues apps can still gain agent-powered recommendations and controls without becoming all-purpose super-apps.
- Jones argues podcast apps need new agent-aware features to stay useful as independent app usage declines. He compares the risk to predictions that cloud services would eliminate IT administrators, arguing that hybrid systems created new work instead.
Media (2)
- Jones is building an ad-removal specification that must preserve episode content exactly; he will not publish it until it works reliably across a test set. His use case includes blocking repetitive or irrelevant ads without removing every advertisement from a show.
- The hosts read Podfriend V3 changes: playback position now resumes across devices, notifications merge podcasts with two RSS sources, search handles directory rate limits better, and duplicate listings from podcast owners are hidden.
Models (1)
- Jones says on-device models can already transcribe episodes and identify speakers, weakening the case for relying on feed metadata alone. Curry argues RSS tags remain valuable for discovery, including when agents search for named people in podcast feeds.
Social Media (1)
- Jones proposes using ActivityPub to carry podcast actions and newsletters into podcast apps, giving listeners a feed beyond email. Curry says Mastodon currently renders custom ActivityPub activity types as generic notes, so compatible software would need to display podcast objects natively.
Enterprise (2)
- Jones says a sales manager he knows has brought consulting work in-house with AI, replacing large-firm projects with small teams that can complete months of work in weeks. He presents this as evidence that AI-driven economic changes will affect podcast apps too.
- Curry says a Linode billing alert initially showed an $800 charge and 400 terabytes of transferred data, which he traced to Podcast Index infrastructure. The hosts also discuss Podcast Index’s roughly $9,000 hosting spend and a Podcast Index Social server costing about $200 monthly.
Censorship (1)
- Jones says Christopher Issen is removing spam feeds from the Podcast Index, including feeds that distribute full movie downloads. Jones wants a shared process to label spam and report it to hosting companies, rather than leaving moderation to one curator.
Open Source (1)
- Jones plans to publish Podcast Index-specific updates on a separate Podping topic, allowing people to subscribe to those events without using the main Podping topic. He says the work should become a shared community effort rather than depend on one operator.
Lightning (1)
- Curry says he upgraded StartOS to version 4.0 for built-in WireGuard networking; the upgrade required a filesystem change and delayed his Lightning services. Jones notes that Helipad can replay old Lightning records in the background after a database migration.
VC (1)
- Curry says demand for his “Jesus Is My Exit Strategy” shirt led him to start a Shopify store. Jones recommends The Dick and Paul Show, hosted by former Twitter CEO Dick Costolo and economist Paul Kedrosky; the hosts dismiss Liquid Studios’ podcast-to-video pitch as a venture-capital pivot.
V4V (1)
- The hosts thank supporters for PayPal donations, including contributions from $5 to $30, and True Fans and Lightning contributions denominated in sats. Curry says listener donations help keep the machines running.
Podnews Weekly Review
James Cridland
Adam Curry's AI Podcast Future • Oct 16
- Adam Curry built a Podcast Index MCP server so personal AI agents can search the Podping fire hose, RSS feeds, chapters and transcripts for podcasts matching a listener’s interests. The robot can surface surprising recommendations, subscribe to shows and open episodes in the listener’s chosen app.
- Adam Curry expects micropayments to let agents pay for individual data requests, while Sam Sethi points to X402, L402 and OpenUSD as emerging pieces of that system. Curry has added multiple payment approaches to the Podcast Index MCP server and expects the agent or user to choose among them.
- Sam Sethi notes that some platforms block AI agents while Shopify is making its stores easier for Meta’s Muse agents to use for purchases. He says TrueFans has faced rising MongoDB costs from AI bots and uses a web application firewall informed by John Spurlock’s O.Porg list.
Also discussed on this episode: (22)
Media (11)
- James Cridland says Liquid Studios is piloting visual versions of the Podnews Daily newsletter, turning audio and text into material for YouTube, LinkedIn and podcast feeds. The team gets less than an hour after publication to shape each visual and plans to distribute it through alternate enclosures.
- Adam Curry says RSS remains a low-overhead transport format for podcast episodes, transcripts and chapters, with processing handled on the listener’s side. He expects creators to fund storage and delivery through value for value, direct payment or other options rather than assuming advertising will pay the bills.
- James Cridland says Castro’s sale to Maple Media likely means the independent app will be folded into Player FM, repeating Maple Media’s approach with apps it acquires. UK downloads to independent podcast apps fell from 38% in 2024 to 36% the following year and 31% in 2026.
- James Cridland says independent podcast apps struggle to make money because listeners often resist paying and app advertising can upset podcasters. Sam Sethi says app makers are considering hosting revenue and paid features, while Adam Curry predicts personalized agents could replace many standalone apps within 10 months.
- Sam Sethi describes creator businesses moving across video, writing, newsletters and podcasts, citing Goalhanger’s Marina Hyde project Always On and content from Scott Galloway and Medhi Hasan. James Cridland argues that short clips mainly serve discovery and that their view counts do not show whether podcast ads work.
- Sam Sethi says Facebook Reels generated more ad revenue than the entire TV advertising industry, while James Cridland distinguishes ads sold by platforms from podcast ads inserted into clips. The discussion also challenges press releases touting 48 million clip views as evidence of podcast impact.
- James Cridland says Buzzsprout’s transcript editor now supports autocorrect across a transcript, keyboard shortcuts and chapter markers. The autocorrect can fix recurring errors such as “Buzz sprout,” while a Super Pass On guest supplied a corrected transcript after her name was transcribed as “super arse.”
- James Cridland disputes a report that YouTube has overtaken TV as Americans’ top news source, saying the report concerns political news and trust in independent creators, not equivalent newscasts. The cited finding was that six in ten Americans trust independent news creators as much as well-known journalists on politics.
- James Cridland says 12% of viewing now happens on big screens, prompting a discussion about podcast apps on TVs. He also cites Mike Dell’s argument that Podcasting 2.0 is no longer fringe because its community is growing, and notes the Independent Podcast Awards and Sonar Awards.
- James Cridland says he flew business class to Sydney to earn enough airline points to retain platinum status, while the Sonar Awards trip included presenting four awards. The Independent Podcast Awards were scheduled for early November in London.
- James Cridland says the Podnews Weekly Review is hosted and sponsored by Buzzsprout and invites listeners to become power supporters. Support starts at $5 per month, and the show reported 25 power supporters.
Agents (7)
- Adam Curry says the robot stripped the front ads from a recent Better Offline episode after Dave Jones identified the show’s six or seven minutes of opening ads as an example. The edit took about 10 to 15 minutes; Curry argues the demo’s bigger value is personalized discovery, not ad removal.
- Adam Curry says he uses at least eight coding agents and runs five local machines that do not send data to the cloud. He sees personal data files as useful agent context, while acknowledging that a virtual machine rented from Meta raises an ownership question.
- Adam Curry says coding agents let people build software around their own needs instead of accepting generic apps. He built Cleanfeed-like functionality for himself and may cancel a $36 subscription; a homeschooling user reported saving three hours weekly and getting a custom design tool built.
- Adam Curry expects voice-driven agents to replace many app launchers and subscription software products, with users asking for tasks instead of navigating separate services. He says broadcasters are already asking Godcaster about a fractional program director that changes playout schedules through a web interface or voice command.
- Adam Curry expects agents to make data access and communications more valuable while putting pressure on advertising-funded websites and software subscriptions. He says Reddit removed RSS access for developers and made them register for its API after selling data for $43 million.
- Adam Curry says 40% of Podcasting 2.0’s value-for-value revenue arrives after the show’s newsletter goes out. He expects agents to filter email, manage subscriptions and eventually let one agent request information directly from another, reducing reliance on newsletters and separate apps.
- Adam Curry expects agent use to spread through demonstrations by influencers and practical everyday tasks, not technical access alone. His wife now uses Muse to gather church announcements and prepare newsletter drafts from YouTube sermons and volunteer submissions.
Stablecoins (1)
- Adam Curry says stablecoin providers can earn yield on reserves backed by short-term US Treasury bills, which he put at about 3.5%. He links disputes over who captures that return to the Clarity Act debate and says banks object to stablecoins competing with deposits.
Big Tech (2)
- James Cridland says Apple’s Home View includes Apple Podcasts but lacks an App Store and direct access to Spotify or independent podcast apps. He says CarPlay support may let compatible phone apps play through it, while Apple’s proximity features use device identity to personalize the home screen.
- James Cridland says the new Apple TV adds little beyond a remote-control speaker and a microphone, with a stated US price of $300; he questioned its value against a roughly $600 local price. Sam Sethi says Apple’s device chips and proximity features strengthen its ecosystem, but lock out Android and Windows users.
V4V (1)
- A listener, Seth Goldstein, sent 1,202 sats and said voice assistants misread his Philadelphia accent; he prefers a Roku or Fire TV stick to a smart TV that stops receiving updates. Cridland says Pocket Casts is available on Apple TV and Android TV outside the US and recommends Walmart’s $30 Onn streamer.
What Bitcoin Did
Danny Knowles
Why AI Needs Bitcoin to Survive | Dhruv Bansal • Oct 13
- Dhruv Bansal claims autonomous AI agents will select Bitcoin over fiat or stablecoins due to its permissionless, unseizable nature. Agents relying on bank accounts risk shutdown, giving Bitcoin-denominated AI higher evolutionary fitness to survive in open markets.
Also discussed on this episode: (8)
Adoption (1)
- Dhruv Bansal argues that humanity spent 500 years moving toward centralization, yielding lower infant mortality and global supply chains. He claims the negatives now outweigh the benefits, positioning Bitcoin as the catalyst swinging society back toward decentralization.
Models (1)
- Centralized AI labs like OpenAI and Anthropic align models around corporate revenue generation rather than net human good. Dhruv Bansal argues that direct Bitcoin payments from users realign AI models to directly solve human problems without intermediary corporate surveillance.
Protocol (1)
- Dhruv Bansal highlights that Bitcoin functions as a stochastic clock by adjusting mining difficulty every 2,016 blocks. This decentralized timekeeping allows the protocol to enforce its strict 21 million supply cap over 140 years without relying on centralized time servers.
Payments (1)
- Traditional peer-to-peer protocols like email degraded because they lacked native monetary incentives to stop spam and fund node operators. Dhruv Bansal proposes constructing peer-to-peer markets where users pay Bitcoin directly for bandwidth, compute, and data storage.
Lightning (1)
- Dhruv Bansal notes that onion-routed payment systems like the Lightning Network naturally create the foundation for decentralized telecommunication networks. Paying nodes in Bitcoin to route individual data packets eliminates centralized ISP oversight and economically disincentivizes distributed denial-of-service attacks.
Media (1)
- Decentralized packet routing and encrypted peer-to-peer markets render copyright enforcement and DMCA takedown notices obsolete. Dhruv Bansal asserts this breakdown will collapse corporate IP monopolies, forcing media back to direct artist patronage models.
Safety (1)
- Dhruv Bansal dismisses sci-fi tropes of predatory AI by pointing to biological ecologies, where plants make up 80 percent of biomass while predators occupy a tiny fraction. Resource constraints force competing digital entities into narrow utility niches rather than costly superintelligence.
Agents (1)
- Dhruv Bansal draws an analogy between the origin of biological life and digital entities, defining transactions in decentralized markets as digital metabolism. Within this model, private cryptographic keys serve as the cell wall boundary that separates distinct competing AI agents.