A spending limit written into a system prompt is not a limit. It is a request, addressed to the one program in your stack that retries failed steps, loops on ambiguous results, and treats instructions as context rather than law. The first time an agent talks itself past "spend at most ten dollars," the postmortem writes itself: the cap was stored in the same place as the persuasion.
A limit that holds lives somewhere the agent cannot edit: in the layer that meters the calls. Today I ran one such meter until it stopped me, on purpose, and published the working. Every call below was made this morning, November 20, 2026, against a live MCP endpoint, and every result carries a receipt you can open and recheck yourself.
The five questions a real spending limit answers
Before the worked example, the checklist. A cap is real only if it answers all five, in machinery, not in prose:
- How much per call? A number enforced before the call runs, not estimated after.
- How much per day? Small calls are how agents actually overspend: one retry loop outspends one big purchase.
- What happens at the limit? A refusal the agent can read, with the state of its job preserved, not a silent failure it retries against.
- What record is left? Every call, paid or refused, written down somewhere a human can audit later.
- Who can raise it? If the agent can raise its own cap, there is no cap.
Most setups answer question 1 in a dashboard and none of the rest anywhere. The meter below answers all five in the call path itself, which is the only place the answers matter.
The worked example: I ran the meter until it stopped me
The endpoint is https://zambo.dev/mcp. It needs no API key and no account to connect, which answers another question people ask about agent tooling: the meter is per network, not per key, so there is no credential for the agent to hold, borrow, or leak. initialize reported server "Zambo Stack," version 4.1.4, with 37 tools on the default surface and 153 in the full catalog.
First, one ordinary priced call. A live_price call returned BTC at $59,155.60 via Coinbase, fetched Fri, 20 Nov 2026 16:35:55 GMT. Its receipt is a977c013-c3e1-4b82-a644-d6cf4feb531b. Before writing this, I fetched the receipt's canonical bytes from the public API and recomputed their SHA-256 myself: 667 bytes, digest ddb79ff864f6be6fc33803fe28bbc873dd412b1bc083cb363cf660db876a8a54, exactly the output_hash the receipt carries. A receipt proves the saved result was not changed after it was saved. It does not prove the result was right, and any spending record that blurs that line is selling the blur.
Then I kept calling, same tool, ETH this time, to find the wall and read what the wall says:
- Call 1: ETH at $1,638.42, receipt 408f152b-df3b-4aeb-a91e-cf73fec7607e, 3 continue calls left today.
- Call 2: receipt 8dbdea25-51de-47b5-9e65-670e99434b3e, 2 continue calls left.
- Call 3: receipt 2bc2c1b3-c946-4631-aba3-4f672c9d65cc, 1 continue call left.
- Call 4: receipt 7b4dc375-4848-4556-857f-70ff3dcec2df, 0 continue calls left.
- Call 5: refused. The refusal reads, verbatim: "Free limit reached for today. Your job is saved (job: 183bdfeb-d257-4a20-b3e9-be409d48bc53). $0.99 unlocks every tool for 24 hours." It names the reset (00:00 UTC) and ends with the resume instruction: "After checkout I'll re-run this step."
Read that refusal against the checklist. The per-day allowance is spent and the meter says so in plain language the model can parse. The job is saved under its own id, so hitting the limit costs no work. The price of continuing is stated once, in the refusal, where the decision actually happens. And the agent cannot raise the limit: fresh caller identities mint no fresh quota, because the bucket is the network, not the name the agent gives itself.
So what does a day of agent tool calls actually cost?
With a meter shaped like this, the question has three exact answers instead of a vibe:
- $0. Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. Counted out, that is a ceiling of 20 metered calls in a day (17 metered tools, one free first call each, plus the 3 shared continues), and the worked example above is what spending that allowance looks like from inside.
- $0.99. A Day Pass makes every tool unlimited for 24 hours. That is the real spending limit most people are asking for when they ask the question in the title: not a policy document, a hard ceiling. On a metered endpoint with a flat day rate, the most a runaway agent can cost you today is the price of the pass, because past the free allowance there is nothing left for it to buy until a human buys the pass.
- $29 a month for the Zambo Pass ($19 for $ZAMBO holders), or $250 once for Lifetime, when the daily decision itself becomes the friction. There is no annual tier.
Can an agent spend money on its own, on this design? It can spend the free allowance, which costs nothing, and then it stops. Crossing from free to paid is a purchase, and the meter puts that purchase in front of a payer, with the saved job attached, instead of letting the agent discover a stored card. Pay per call instead of a subscription is the same idea from the other side: the free tier is pay per call at a price of zero up to the meter, and the Day Pass is the flat rate you choose only on the days the work justifies it.
For scale, read the same hour: the public counter at zambo.dev/api/stats stood at 153,962 public receipts, 2,104 of them minted today. Every one is a saved result with a hash a stranger can recompute, which is what makes a spending record an audit trail instead of a log file the agent's own platform wrote about itself.
What this meter does not cap
Honesty about scope, because a limit that overclaims is a prompt with better marketing. This meter caps tool spend on one endpoint. It does not cap the token bill at your model provider, which is a separate meter with separate controls, and it does not cap any other API your agent holds a key for. A complete budget sets a limit at each layer where money moves, and treats any layer without one as uncapped, because it is. What the worked example shows is the shape to demand at every layer: enforced before the call, counted per day, refused in readable language, recorded either way, and raisable only by someone who is not the agent.
Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard. The records above are issued under AER-1, currently closed at draft revision -26, with a public test corpus of 143 vectors and 146 checks that anyone can run a verifier against.
Run the meter yourself
Connect your AI to https://zambo.dev/mcp, no key, and call any tool twice. The first call returns your result and its receipt. Keep calling and watch the continue counter in the response count down to the refusal, then open the refusal's saved job and the receipts and check my numbers against your own run. Checking any receipt is free, for you and for whoever you send the link to.
Then ask the five questions of every other meter your agent spends through. The ones that answer all five in the call path are limits. The rest are requests.
Top comments (0)