Prince Maurya · Agent reliability
Audit / 2026
Know what breaks before your users do.
A fixed-scope audit of one agent surface. I reproduce the failures at volume, trace each one to a cause, and leave you with a written report and a regression test — so the next silent failure shows up in CI, not in production.
Replies within 24 hours · [email protected]
01 — What you get
Three answers, in order.
Most agent problems are not “the model is bad.” They are a prompt assumption, a tool contract, or a missing gate. The audit separates those three.
Reproduce
I run your agent on the real tasks, not the demo task, and I run each one repeatedly. A failure that appears on run 3, 7, or 20 is the whole point — one clean pass proves nothing.
Trace
Every reproduced failure gets a trace: the input, the tool call, the tool response, the reasoning step where it went wrong, and the cause. You get the evidence, not a guess.
Gate
The audit ends with a regression test that fails on the bug you had, plus a short gate you can run in CI. The same failure cannot silently return.
02 — Evidence
Why repetition is the method.
This is measured on my own public work, not on a client engagement. The audit exists because this drop is normal, and most teams only ever measure the first number.
Measured on nano-agentrl · public repo, public numbers
28%
4%
Same model, same task, eight consecutive runs: 28% on a single pass, 4% across all eight. The gap is what an audit looks for — not whether the agent can work, but whether it keeps working. Read the measurement ↗
03 — Process
Five business days, start to report.
You stay in the loop with short written updates. There is no weekly call to sit through, and no dependency on your API keys.
You send access
A read-only repo or a zip, the task the agent is supposed to do, and one example of a failure you have seen. I reply with two clarifying questions and a start date.
I reproduce and trace
At least 20 runs on the surface, traces on every failure, and a ranked list of causes by user impact. If nothing reproduces after two days, I tell you immediately and we stop.
Report and gate
You get the written report, the run logs, the regression test, and a 30-minute walkthrough. Everything is yours to keep and change.
04 — Scope
Fixed scope, so the price can be fixed.
One surface keeps the audit honest and the timeline real. If you need more than this, the next step is a second audit, not a bigger invoice.
Included
- One agent surface — one entry point, one task family
- Python or TypeScript, up to five tools
- 20+ runs with saved traces and logs
- Written report, ranked by user impact
- One regression test and a CI gate
- 30-minute walkthrough
Not included
- Building features or rewriting the agent
- Multiple agents or surfaces
- Ongoing monitoring after the report
- Model training or fine-tuning
- Hosting, deployment, or on-call
- Work without read-only access to the code
One fixed price. No retainer, no usage fees.
Invoiced on delivery. If nothing reproduces in the first two days, I send what I tried and there is no invoice — reproducibility is the product.
01What access do you need?
Read-only repo access or a zip, plus a short description of the task the agent is meant to do. I do not need production credentials or API keys — I run against my own keys or a sandbox.
02Is the work confidential?
Yes. I will sign your NDA. Findings are yours, and nothing from the audit is published or used as a case study without your written permission.
03What if nothing reproduces?
You get the run matrix and a note on what I tried within the first two days, and there is no invoice. An audit that invents findings is worse than no audit.
04Who is this for?
Teams with an agent already in use, or about to ship, where failures are intermittent and the cause is not obvious. If you are still choosing a framework, this is not the right first step — start with the writing.
Start here
Five days. One surface. A clear answer.
Send the repo and the task list. You'll get a reply within 24 hours with two questions and a start date.