Your agent spent it

Hey there 👋

An agent can hold a credit card and spend from it. That stopped being the interesting part about a month ago. The question I keep turning over is the ugly one underneath it: when the agent buys the wrong thing, who eats the loss?

Here is my bias, and I will put money on it. The banks slamming the door on agents are the adults in this story. They are the only players asking the question that decides everything, and they are answering with a flat no while everyone else ships demos. Getting an agent to pay is a solved problem. The open question is who covers the bill when it buys garbage, and right now the honest answer is nobody.

Three stories today, one argument. An agent that can spend is easy. An agent you would hand your card to is the whole game.

- Ron

The Big Thing

Agents started spending real money, and almost nobody trusts them to

Something shifted in September. Stripe president John Collison said agentic commerce "started going vertical," his words for a curve that bent up fast. Stripe is not putting hard numbers on the record, so treat the shape as real and the size as unconfirmed.

The texture backs him up. Dan Hill, who runs the consumer side of Stripe Link, said Link now powers about 200 AI agents, up from roughly 70 a month earlier. His read: "the volume we're seeing now is becoming very material." The named agents are a who's who of the moment, including Instinct, Meta Muse, xAI's Grok Bot, Skydive, Town, and Underdog. Anthropic and OpenAI use Link too, though not yet for agent purchases.

You will see louder figures floating around, growth multiples and average-spend numbers. They come from a paywalled report, and nobody outside Stripe has checked them, so we are leaving them out. The climb itself is on the record, but the precise size is not something we can confirm from here, so treat it as a question mark.

Now the part that should stop you. In a PYMNTS Intelligence survey from September, 93% of merchants said the company behind the agent should eat the losses when it buys wrong. A separate Product.ai survey in April found only about 14% of consumers trust an agent to finish a purchase without a human checking it first. So money is moving at real volume, and almost no one will stand behind it.

That gap is the whole story. Every other item today is one more attempt to answer a single question: who is on the hook when the agent is wrong? Block it, insure it, or prove the human signed off.

If you are building agents for B2B, the spend path is now your gating design decision, ahead of which model you pick. Who authorizes a purchase. What the hard cap is. Who covers a bad buy. Write those three down before you open a model picker.

Why it matters: the money is already moving, and the part that decides who pays for a bad buy barely exists yet. Ship it? WATCH. The rails are moving faster than the rules, and the first vendor to make "we cover the mistake" credible will pull the market toward it.

Sources: Upstarts Media on how AI assistants spend through Stripe Link and Forkast on agentic commerce going vertical.

Tour de Headlines

🏦 Schwab and JPMorgan are quietly blocking agents from moving money. Charles Schwab CEO Rick Wurster gets a weekly report of blocked agent attempts, and he says it runs in the "tens and twenties." The firm stops unauthorized agents, naming Meta Muse, from acting on client accounts. "We don't want someone just giving their password to Muse, and letting Muse come in and take action on that client's behalf," he said. Jamie Dimon confirmed JPMorgan blocks outside agents from reaching customer data too. What stops agents here comes down to permission. The people who hold the accounts decide whether an agent can act, and so far the answer is no. Bloomberg.

🪪 World puts an identity suite in front of the banks. At Money20/20 on Sunday, October 18, Sam Altman's World launched its first ID suite built for banks and financial firms. It bundles World ID, World ID for Zoom, and World ID for Agents (AgentKit). The part that matters is delegation: you can hand your verified human "Proofs" to an agent, which presents them to third-party platforms. That is the primitive the banks keep blocking on, proof a real person signed off. Design partners include OpenExchange, Bain Capital, and VanEck, though no bank is named as a live user yet. World's cited Deloitte figure of $40B in AI fraud by 2027 is Deloitte's number passed along by World, not one we traced. Early, but aimed at the right problem. The Block.

💳 Musk says SpaceX will make you whole if Grok buys wrong. Here is the fun one. After a demo where Grok Bot booked a flight from a photo of a credit card, Elon Musk said SpaceX will reimburse users, "make you whole," if the bot mishandles a purchase or leaks card data. He did not say what the limits are, so do not read it as unlimited. Still, it is the first vendor saying the quiet part out loud: we will eat the mistake. That is the exact trust mechanism 93% of merchants asked for. Expect every agent buyer to start demanding the same promise in writing. The Next Web.

Tool of the Day

🛠️ Harness by Autonomous

One desktop cockpit for running several coding agents at once without them stepping on each other.

If you juggle Claude Code, Codex, Hermes, and Cursor, Harness puts them in a single window on Mac or Linux. Each agent gets its own Git branch, so they do not clobber one another's work, and you can watch all of them from your phone over an encrypted link. Autonomous lists 49 domain harnesses, including CAD and robotics. One caveat: we have not run it, and the feature list is the vendor's, from its Product Hunt launch. The team points to an "openharness" repo, but we could not confirm the license, so we are calling it a free desktop app until that checks out. Verdict: TRY if you run more than one agent and keep losing track of which did what.

See Harness on Product Hunt

Worth a Click

• 🏛️ Agents are already working at city hall. Virginia used agentic AI to scan regulatory documents and, as reported by the Urban Institute, turned up more than $1.4B in annual savings while cutting permit times by around 80%. Utah runs autonomous prescription renewals with Doctronic, where physicians agreed with 91% of the AI's calls, again per the Urban Institute writeup. Boston wired agents into its open-data portal through an Anthropic MCP server. Treat the figures as reported numbers we have not audited.
• 💰 AIsa gives agents a wallet and a 5,000-API marketplace, and Tether just backed it. Tether Capital invested an undisclosed amount in AIsa, which lets agents find, call, and pay for APIs, skills, and models one call at a time, with user-set spend caps. The company reports more than 250,000 agents and over 100M cumulative transactions. Those counts are AIsa's own, so weigh them accordingly.
• 🌲 A new RL recipe makes research agents measurably better. TreeOPSD, from Tencent Hunyuan, hands out per-turn credit from sparse rewards using branching rollouts. On Qwen3-30B it raises average accuracy by 4.97 points over GRPO on deep-research tasks. It is a preprint, not peer-reviewed, so read it as a promising result and nothing settled yet.

Build the spend path before you pick the model. Everyone this week answered the same question three ways: Schwab blocks it, Musk insures it, World proves it. For your own agent, write down who authorizes a buy, what the hard cap is, and who covers a wrong one. Do that before you argue about which model is smartest. That is the call that ships or sinks agentic commerce in B2B.

- Ron