When an agent loops overnight or reaches a system it should not, the bill and the questions arrive together. Abacross puts the spend limits, access boundaries and tamper-evident logs in place on AWS before that happens, so the answer is a record, not a reconstruction.
Fixed fee, quoted before any work starts. No per-token billing.
$ ajv verify journal \ --require rfc3161,opentimestamps records 41 [ok] digests 41 of 41 reproduce [ok] chain links every record links to its predecessor [ok] sequence continuous ... [ok] block 40 root 64db613e3dc7 anchored by opentimestamps + rfc3161 $ echo $? 0
Our own journal: 41 records as of 24 September 2026. Verify it yourself.
They are the same three every time, and all three are infrastructure problems rather than model problems.
Budget alerts arrive after the money is gone. We put circuit breakers in the execution path, so a recursive call stops at a limit instead of at an invoice.
Credentials get shared, scopes get widened for a demo, and nobody narrows them again. We set access boundaries that survive the next deadline.
After an incident, ordinary logs cannot show they were not edited. We build provenance you can hand to somebody who does not trust you.
You can stop after any of them, and the first one is read-only.
It starts with the free readiness check, run by you. Then read-only access and a walk through what the agents actually do. You get a written findings list, ranked, with the blast radius of each one spelled out. Nothing changes during this step.
We implement the findings you choose, as reviewable infrastructure-as-code. Spend limits, access boundaries, and a provenance trail, each one a diff you can read before it is applied.
Runbooks, the reasoning behind each decision, and the tooling to re-run the checks yourself. The goal is that you do not need us again for the same problem.
Why an agent's record has to be kept by someone who is not the agent: the diary, the hash chain, the anchor, and how anyone can check it. The verifier and our own journal are open on GitHub.
The clearest evidence of how we work is what we have already worked out. No gated whitepapers and no email wall.
A role that is not an administrator can become one, and no rule has to be broken on the way. This post is the prose version of our ninety-second film...
An agent works through the night. In the morning the bill is up, a bucket is gone, and the first question anyone asks is the one this post is about...
Corrected on 2026-09-24. The first version said a budget alert and a budget action fire within minutes of the line being crossed. They do not: AWS...
Read everything ›
Browser-based utilities that never send what you paste anywhere. They are here because they are useful, not because they are a funnel.
One read-only file, twenty API calls, three questions: can an agent here spend, reach, or act without a record? Scores your account and prices the fix.
The full range of outcomes across thousands of market scenarios, not one optimistic number.
Check an Agent Card's fields, decode its JWS header, and see what a signature does and does not prove.
Paste one log event or span to see which provenance fields you would wish you had after an incident.
Tell us what you are running. The first reply will be substantive rather than a calendar link.
Request an audit
Opens your email client with a short prompt already filled in. No form, no tracking, nothing stored here.