I map who decides inside your agent system
Your agent system makes a lot of small decisions over and over. May this tool run? Is this answer good enough? Is the task finished? Some are held by code, some by a model, some by a person. Some are held by nobody. Most teams have never seen the list.
I build that list. I have done it in public for twelve open-source agent systems, and each map found between 85 and 279 recurring decisions. In OpenHands, the map turned up a completion check that read “false” as complete. I reproduced it with a failing test and reported it upstream.
What you get
- A map of the recurring decisions found in your system: the question, who holds it, and, where found, how it has failed and what checks it.
- A gap report: decisions nobody holds, checks that cannot fail, and places where a person is assumed but never asked. Each finding names the file and line.
- The worst problems reproduced with a failing test, and fixes as pull requests if you want them.
- A page your colleagues outside engineering can read.
How it works
- You give me read-only access to the repository. It takes about two weeks. The price is fixed: $10,000 for one system.
- A second model family checks every sentence about your code against the code before you see it.
- If you want the map kept current as the system changes, upkeep is $2,500 a month.
Who it fits
Teams with an agent system in production, or about to be, and someone who can take a pull request and argue a finding.
Write to me through the contact form. I answer every one myself.