Everyone is watching what your model costs. Nobody is watching what your agent spends.
GetAgentCap team · published 9/20/2026
Look at the AI cost tooling landscape and you'll notice almost all of it is pointed at one number: tokens in, tokens out, dollars per model per day. That's a real problem and there are good tools solving it. But there's a second budget sitting right next to it that almost nothing is watching — the money the agent itself spends downstream, paying for the things it decided it needed. Those are different budgets with different failure modes, and only one of them has an industry.
Two budgets, one of them unwatched
Budget one is inference. You call a model, you're billed by a provider you have an account with, on a bill you can see, with a dashboard, usage limits, and a hard stop if you want one. There's an entire category of tools making that number smaller and more legible, and they're good at it.
Budget two is whatever the agent buys. A paid API behind x402. A data feed. Compute. Another agent's service. These payments are made from a wallet, in USDC, settled on chain, with no account relationship, no invoice, no monthly statement, and no provider-side limit — because the entire premise of machine-to-machine payment is that no account relationship is needed. That's the feature. It's also why the spend is invisible by default.
Here's what makes the second one meaner: budget one is bounded by your own code path. You know roughly how many model calls your system makes, because you wrote the calls. Budget two is bounded by the agent's judgment about what's worth buying, which is exactly the part you delegated on purpose.
Why the LLM-cost tools can't see it
This isn't a gap those tools are neglecting out of laziness — they're instrumented in the wrong place. An LLM observability layer wraps your model client. It sees prompts, completions, latency, token counts. It sees the agent deciding to buy something, if that decision even surfaces as a tool call. It does not see the settlement, because the settlement happens on a chain the SDK has no reason to know exists.
The place where budget two is legible is the chain itself. A USDC Transfer event is unambiguous, timestamped, final, and can't be lost by a middleware that wasn't wrapped. You don't have to trust the agent's own account of what it spent, which matters more than it sounds like it does — an agent's self-reported spend is a log line, and log lines are written by the same code that's misbehaving.
So we read the chain. Public address in, transfer events out, attribution by wallet. It's a narrower input than an SDK integration and that's the trade: we know less about *why* a payment happened, and we can't be wrong about *whether* it happened.
Where we deliberately stop
We don't enforce. We don't hold keys, we don't sign, we're not in the request path, and we cannot block a payment. There are tools that do sit in that path and act as a policy layer, and if you need a hard ceiling that physically cannot be exceeded, one of them is your answer, not us. That's not false modesty — it's a real fork in the road and you should know which side you're on before you pick anything.
What we're betting on is that most fleets right now don't have an enforcement problem, they have a knowledge problem. Nobody in the stories I keep reading said "I had a budget and it failed to hold." They said they found out from the balance. The gap between money leaving and a human knowing about it is the thing we're trying to close, and closing it requires exactly one piece of information from you: an address you can paste from a block explorer. No keys, no wallet-connect, no SDK, no code change.
And no, we're not the only people who've noticed agent spend exists. There are a handful of x402 explorers and a few funded teams circling the enforcement side. What we haven't found is someone treating downstream agent spend as a monitoring problem for small teams rather than an enterprise FinOps product or a block explorer. That's a narrower claim than "nobody's doing this," and it has the advantage of being true.
What this means for your fleet
Go check whether you can answer this right now, without opening a block explorer: how much USDC did each of your agents spend last week? If you can't, you have an unwatched budget, and its size is bounded by your wallet balance rather than by anything you decided.
If the answer is "almost nothing," great — that's the correct answer for most fleets today and you should spend your attention on inference cost instead. The reason I built this is that the answer stops being "almost nothing" the moment an agent's loop includes paying for things, and the transition from trivial to alarming takes about one bad afternoon.
Two budgets. Watch both. The second one is the one with no invoice coming.