What changed
A classification that repeats
The classification gives the same result from the same data every time, and a benchmark checks it.
Agents under shared rules
Defined the product, the code structure and the agents’ shared rules, and checked quality. Agents wrote most of the code.
1,410 code changes, June–October 2026
Priorities with reasons
The classification shows which services and pages searches relate to, so what gets fixed first can be justified and checked afterwards.
Problem
A large healthcare website has a lot of search data. Classified by hand, the result depends on who does it and when, and priorities cannot be justified afterwards.
Value
The classification shows which services and pages searches relate to. The same data always gives the same result, so what gets fixed first can be justified and checked afterwards.
Role
Product definition, code structure, classification, benchmark tests and quality checks. Agents wrote most of the code.
- Defined a classification that gives the same result from the same data every time, plus a benchmark for it.
- Wrote the shared agent rules and set what the workbench and the other parts each do.
- Kept track of release versions, design decisions and audits.
Limits
- Employer-related system, described only in general terms. Code and material stay private.
Sources
- 1,410 code changes across branches, June–October 2026
- Search visibility: first version baseline
- WP-03, repeatable benchmark