Michal Cichra is a principal engineer at Safe Intelligence working on Spec27, a platform for validating AI agents. He builds systems that preserve architectural decisions, enforce engineering standards, and keep agent-generated software testable after its original instructions disappear.
Cichra joined API-management company 3scale in 2011, developing distributed infrastructure for APItools, business-data systems, and a substantial Ruby on Rails product. After Red Hat acquired 3scale, he became a principal engineer, helping bring its hosted platform to OpenShift and release its software as open source.
His projects included APIcast API Gateway and 3scale Zync, which synchronizes application data with external identity providers. Red Hat credited him with creating Zync; he also developed open-source Lua utilities for API infrastructure. He subsequently worked at Microsoft before joining Safe Intelligence, where he contributed to Spec27’s launch and applied his experience maintaining long-lived products to autonomous software development.
Making engineering decisions enforceable
- Architecture decision records as operational memory: Capture why a rule exists and connect it to automated enforcement. Import restrictions can prevent rendering code from accessing databases, while returning plain data instead of ORM objects reduces opportunities for N+1 queries. Failed checks direct agents back to the decision record so they can repair violations with the architectural rationale intact.
- Executable behavior-driven specifications: Connect lightweight product requirements and critical user journeys to readable Cucumber scenarios that engineers can review and automated systems can execute. End-to-end tests remain isolated from the database, forcing validation through actual user-facing behavior.
- An enforceable agent feedback loop: Combine Git hooks, linters, type checks, architectural constraints, and continuous integration into the same mandatory verification cycle. Adapt that cycle to individual tasks: browser-based iteration for interface changes, targeted tests for backend modifications, and reusable design systems for visual consistency.
Cichra’s AI Engineer Europe session connects these techniques to a practical goal: coding agents capable of working autonomously for extended periods while recovering team decisions after their context is compressed.
