We spent a week working through a large code base refactoring with around 500 agents. The learnings from that work went directly into the ArchKeel releases up to 0.8.3.
We still need to understand which components exist, what they are responsible for, and whether the changes and new code stick to our rules. More agents make building faster and harder to stay on track.
So we worked on the checks and the reports together. Recursive contracts, evidence for individual rules, and much better exploration of the target and actual architecture. Open a component, inspect its children, follow dependencies, see what was checked and what remains UNKNOWN.
This is what I mean by deterministic harness engineering: explicit architecture decisions, repeatable checks against the code, and missing evidence that stays visible.
Choosing the right architecture still needs judgment. I want the harness to give me something concrete to judge.
ArchKeel is open source, MIT licensed. You can explore a real report here: https://archkeel.com/#explore
Comments URL: [https://news.ycombinator.com/item?id=49926046](https://news.ycombinator.com/item?id=49926046)
Points: 2