Your agents work in the demo. This finds out why they don't work on Tuesday.
I review an agent system you already run, in production or close to it, and hand back a ranked list of what will break, why, and the smallest fix for each. Fixed scope. Written report. You keep the findings whether or not we work together after.
Teams running Claude Code, Codex, Cursor, or custom agents against real repos, inboxes, or customers.
Solo builders with a growing pile of skills, prompts, and subagents that has started to drift.
Anyone who has had an agent do something confidently wrong and could not say why.
01 Instructions and skills
Do the descriptions actually trigger? Are there stale facts, overlapping skills, rules stated in prose that should be a test or a hook? The same five passes as the Skill Corpus Kit curator, done by hand and in depth.
02 Write paths and approvals
Which agents can send, post, merge, spend, or delete. Whether a human approves those actions, and whether that approval can be skipped.
03 Secrets and blast radius
Where credentials live, which agents can reach them, and what one bad run could touch.
04 Memory and context
What the agent remembers, what it forgets, and where old context quietly overrides new facts.
05 Failure handling
What happens on a timeout, a rate limit, a half-finished run, or a bad model output. Whether anyone finds out.
06 Cost and model routing
Whether expensive models are doing cheap work, and where a smaller model or no model would do.
A written report. Every finding is ranked by severity (what it could cost you), not by how interesting it is.
For each finding: the evidence, the risk in plain words, and the smallest fix.
A short list of checks to automate so the same problems cannot come back.
A call to walk through the report.
Not a rewrite. I do not change your code or accounts during the audit. Read-only access, or you share exports.
Not a penetration test or a compliance certification.
No upsell built into the findings. If a fix is a one-line change, the report says so.
01
A short conversation to agree on scope.
02
You grant read-only access (repo, skill folders, agent configs, sample logs) or send exports.
03
You get the written report.
04
A walkthrough call. Any follow-on work is a separate conversation.
The contact form is enough. These are the parts that make the first note useful.
01
What agents or tools are you running, and on which models?
02
What can they write to or send without a human approving it?
03
Roughly how many skills, prompts, or subagent files are there?
04
What is the last thing an agent got wrong that cost you time or trust?
05
Where do credentials for those agents live?
06
What would "this audit was worth it" look like for you?
If this is the review you want, write and tell me what you are running.
victorhustad@victorcollective.comCompanyCam case study.
Autonomous agent system, in the case studies.
Research writeup.
The skill review from this audit, in a form you can run yourself.