From vibe coding to harness engineering: modernizing a regulated environment with AI
From vibe coding to harness engineering: modernizing a regulated environment with AI by Anyul Rivas
Room 5 · Conference · 09:30-10:20
Summary
Section titled “Summary”AI-enabled migration of a banking mainframe to the cloud required a shift from “vibe coding” to “harness engineering.” The talk shows how to control AI-generated code in regulated environments through spec-driven development, evolutionary pipelines, and continuous validation, balancing agility, safety, and changing security rules.
Key points ⭐
Section titled “Key points ⭐”- 22:00 Review ovoid re-deriving context
- 31:00 Review this part to get the slide
Timeline
Section titled “Timeline”- challenge


- is it just a POC or something Real ?


- grill me with Docs
- Verify The plan is easier than very[?] 300 files generated
- No SDD But a custom plan
- Archify skill ?
- Vibe coding -> Demo -> create harness for production Redy
- vibe code To validate the plan an Demo to Business
- But No vibe coding for production


- Demo to production.





- Better to say what to do instead of No To do
- Token Reduction + AI Fin ops

- Review ovoid re-deriving context



- sub agent is controversal
- => Not working
- Human => computer [?] same problem

- TDD Didn’t work
- Anthropomorphize[?] Human process






- what is our job Now
- Review this part to get the slide





- Harness Score card
- https://github.com/anyulled/ai-harness-scorecard





Extra notes
Section titled “Extra notes”- slide
- https://docs.google.com/presentation/d/1qY78jdOwsyRLk-1MmpQPQ1v5pktPMuyOVWKNeFhN74Y/mobilepresent?slide=id.p
- The guy told me there are 20,000 credits per person and that it could go up to 150,000 credits. At the start they blew their quota, even the company quota, so they had to rework that to improve their token usage. For vibe coding they used everything like Claude Code, Cursor, etc. Then, when they entered the production phase, they used tools like GitHub Copilot.
- For sharing skills etc., they didn’t put them in the repo because they couldn’t (I don’t really know why), so they shared them with each other manually.