Your next marketing agent needs a wallet policy, not a bigger prompt library

Circle just announced the Circle Agent Stack.

Agent Wallets. Agent Marketplace. Nanopayments. A CLI for developers. Agents that can hold USDC, discover services, transact with other agents, and settle payments automatically.

They say x402, their HTTP payment protocol for AI, processed $24 million in the prior 30 days. 99.8% settled in USDC.

That is not an experiment. That is infrastructure being used for real transactions at scale.

I have been watching the AI conversation for two years. Most of it has been about what agents can say. This is the first week it felt clearly about what they can pay for.

And that changes the question businesses should be asking.

Not “which AI tool should I use?” Not “how do I get a better chatbot?” The question is: if my marketing agent can spend money, what is the policy around that?

Because here is the thing. The wallet is not the interesting bit. The policy around the wallet is.

An agent without a spending policy is just automation with a credit card. Useful in a demo. Dangerous in practice.

Think about what a marketing agent in 2026 might actually need to spend money on. API costs to fetch real-time pricing data. An AI image generation call to produce an ad variant. A data enrichment lookup on a prospect. A tool subscription to process a segment. An A/B test on a paid channel. A media credit on a platform that now has an agent-accessible buying layer.

Some of those decisions should happen autonomously. Some should not.

That is the conversation nobody is having clearly enough yet.

Most businesses are still stuck on two questions: which AI tools should we use, and who owns the AI strategy. Both are reasonable. Neither touches the operational reality of what happens when your agent actually touches your systems, your money, and your customers.

The operational reality is this. Agents that can act need rules about when to act, what to spend, what to log, and when to stop and ask a human.

At Foundry we have started calling this the wallet policy stack, and it is much simpler than it sounds.

It starts with the job. Not the agent’s general capability, but its specific job on a specific workflow for a specific outcome. Then the minimum tools and data it needs to do that job, nothing more. Then a spending threshold. For a marketing ops agent, something like twenty to fifty pounds before a human review kicks in. That covers API costs and small test budgets without letting an agent rogue-spend on a failing campaign overnight.

Review triggers sit on top of that. Anything that touches a client externally. Anything above the budget threshold. Anything that creates a public artefact. Anything that depends on context the agent might not have recently updated. And underneath everything, a log. What it did, what it spent, what changed, and why. Not for compliance theatre. For when something breaks.

Because something will break.

The companies that build this infrastructure early are going to look boring for a while. Then they are going to be the only ones with agents that are actually trusted with real work. The ones that skip straight to “fully autonomous AI employee” are going to have an incident, get nervous, and either lock everything down or give up on agent workflows entirely. Neither outcome is useful.

And this connects to something I care about beyond the operational efficiency argument.

If agents are going to act as economic actors, they should also carry accountability. At Foundry, and through the infrastructure we are building around Zenko, the ambition is not just that agents fund their own operating costs. It is that they have proof obligations. A record of what they spent, what they created, what they funded. Including a slice of verified good that does not disappear into a sustainability PDF nobody reads.

That is not ready as a mainstream product yet. But the conversation Circle started this week makes it more concrete, not less.

An agent wallet without policy is a liability.

An agent wallet with policy, logs, approvals, and proof is the beginning of a useful operating unit.

The companies building that now are three to five years ahead of where the rest of the market will be when it finally accepts that the prompt library is not the moat.

The policy is.

---