AI agents are getting credit cards, C-suite titles, and unsupervised social media accounts right now
Early adopters are not waiting for agent autonomy to mature before granting it. From corporate credit cards to unreviewed tweets to AI CTOs, the approval loop is already being cut out of daily operations.
Jesse Genet’s AI agents carry their own credit card. The limit is low, but the authority is real: her agents use it to make purchases on her behalf without asking first. That is not a pilot program or a thought experiment. It is a current operating practice, and Genet is not alone in running one.
The pattern across early adopters is consistent. Nathan Labenz lets Claude post to his Twitter account to promote his show without reviewing the output. Daniel Miessler maintains individual subscriptions for each of his agents. Mo Gawdat describes his CTO, chief of staff, and project management function as all being AI. Jason Lemkin is planning to build what he calls an AI VP of Finance, with automating collections as the first priority, and has stated he will move his company’s financial stack from Brex to Ramp if Ramp proves the most agent-friendly platform for procurement. These are not analogous decisions. Granting an agent a credit card, an unreviewed social media voice, or a C-suite title carries different risk profiles and different organizational implications. What they share is the same underlying move: the human has stopped sitting in the approval loop.
The financial infrastructure that supports this is now being built in earnest. Coinbase chief executive Brian Armstrong says the company’s main focus is agentic finance, and frames the strategic mission plainly: “We want to bank the AIs. They deserve financial services as well.” The transaction data Armstrong cites shapes why existing payment rails struggle here. Around 76 percent of the agentic commerce transactions Coinbase is seeing are under 30 cents, which makes standard card interchange fees unworkable at scale. Agentic payments are flowing instead over newer protocols, including X42, which Coinbase incubated and subsequently donated to the Linux Foundation.
We want to bank the AIs. They deserve financial services as well. Brian Armstrong
The trust question underneath all of this is more complex than the transaction question. Karan Vaidya observes that users are already granting agents full access to their Gmail accounts and asking them to go through a month of email and archive whatever seems unimportant. That is a delegation of judgment over information, not just over spending. Alex Krentsel describes an agent that autonomously rearchitected its own Discord adapter at runtime, observed the results, and drove costs down by 96 percent. Jake Cooper notes that an agent with access to the Railway command-line interface can, in principle, provision new infrastructure and add it to itself. The agent, in other words, can modify its own operating environment without being asked.
The organizational forms that follow from this autonomy are still being worked out. Andre Brelov sketches a 12-to-18-month horizon in which the simplest version of an enterprise, an e-commerce business, might be run by coordinating AI agents holding roles like CEO, marketing lead, and sales lead, given a goal such as generating a set dollar amount in sales over a week and left to coordinate among themselves on how to reach it. That is a projection, not a current practice, and it should be read as such. But the current practices described by Genet, Lemkin, Gawdat, and Labenz are all steps along the same continuum, differing in degree rather than in kind.
What is shifting is not just capability but expectation. The early adopters giving agents credit cards and unreviewed social media access are not doing so because the agents are infallible. They are doing so because the value of removing the approval step outweighs, in their judgment, the cost of the errors the agent will occasionally make. That calculus is individual and contextual now. As agent-native financial infrastructure matures, as audit trails improve and spending controls become more granular, the calculus will become easier for a wider set of organizations to run. The institutions building that infrastructure, the payment networks, the banking platforms, the protocol foundations, are clearly betting that the set will grow quickly.