BONUS Transactional AI Development - Commit, Validate, and Rollback With Sergey Sergyenko
AI Assisted Coding: Treating AI Like a Junior Engineer - Onboarding Practices for AI Collaboration In this special episode, Sergey Sergyenko, CEO of Cybergizer, shares his practical framework for AI-assisted development built on transactional models, Git workflows, and architectural conventions. He explains why treating AI like a junior engineer, keeping commits atomic, and maintaining rollback strategies creates production-ready code rather than just prototypes. Vibecoding: An Automation Design Instrument "I would define Vibecoding as an automation design instrument. It's not a tool that can deliver end-to-end solution, but it's like a perfect set of helping hands for a person who knows what they need to do." Sergey positions vibecoding clearly: it's not magic, it's an automation design tool. The person using it must know what they need to accomplish—AI provides the helping hands to execute that vision faster. This framing sets expectations appropriately: AI speeds up development significantly, but it's not a silver bullet that works without guidance. The more you practice vibecoding, the better you understand its boundaries. Sergey's definition places vibecoding in the evolution of development tools: from scaffolding to co-pilots to agentic coding to vibecoding. Each step increases automation, but the human architect remains essential for providing direction, context, and validation. Pair Programming with the Machine "If you treat AI as a junior engineer, it's very easy to adopt it. Ah, okay, maybe we just use the old traditions, how we onboard juniors to the team, and let AI follow this step." One of Sergey's most practical insights is treating AI like a junior engineer joining your team. This mental model immediately clarifies roles and expectations. You wouldn't let a junior architect your system or write all your tests—so why let AI? Instead, apply existing onboarding practices: pair programming, code reviews, test-driven development, architectural guidance. This approach leverages Extreme Programming practices that have worked for decades. The junior engineer analogy helps teams understand that AI needs mentorship, clear requirements, and frequent validation. Just as you'd provide a junior with frameworks and conventions to follow, you constrain AI with established architectural patterns and framework conventions like Ruby on Rails. The Transactional Model: Atomic Commits and Rollback "When you're working with AI, the more atomic commits it delivers, more easy for you to kind of guide and navigate it through the process of development." Sergey's transactional approach transforms how developers work with AI. Instead of iterating endlessly when something goes wrong, commit frequently with atomic changes, then rollback and restart if validation fails. Each commit should be small, independent, and complete—like a feature flag you can toggle. The commit message includes the prompt sequence used to generate the code and rollback instructions. This approach makes the Git repository the context manager, not just the AI's memory. When you need to guide AI, you can reference specific commits and their context. This mirrors trunk-based development practices where teams commit directly to master with small, verified changes. The cost of rollback stays minimal because changes are atomic, making this strategy far more efficient than trying to fix broken implementations through iteration. Context Management: The Weak Point and the Solution "Managing context and keeping context is one of the weak points of today's coding agents, therefore we need to be very mindful in how we manage that context for the agent." Context management challenges current AI coding tools—they forget, lose thread, or misinterpret requirements over long sessions. Sergey's solution is embedding context within the commit history itself. Each commit links back to the specific reasoning behind that code: why it was accepted, what iterations it took, and how to undo it if needed. This creates a persistent context trail that survives beyond individual AI sessions. When starting new features, developers can reference previous commits and their context to guide the AI. The transactional model doesn't just provide rollback capability—it creates institutional memory that makes AI progressively more effective as the codebase grows. TDD 2.0: Humans Write Tests, AI Writes Code "I would never allow AI to write the test. I would do it by myself. Still, it can write the code." Sergey is adamant about roles: humans write tests, AI writes implementation code. This inverts traditional TDD slightly—instead of developers writing tests then code, they write tests and AI writes the code to pass them. Tests become executable requirements and prompts. This provides essential guardrails: AI can iterate on implementation until tests pass, but it can't redefine what "passing" means. The tests represent domain knowledge, business requirements, and validation criteria that only humans should control. Sergey envisions multi-agent systems where one agent writes code while another validates with tests, but critically, humans author the original test suite. This TDD 2.0 framework (a talk Sergey gave at the Global Agile Summit) creates a verification mechanism that prevents the biggest anti-pattern: coding without proper validation. The Two Cardinal Rules: Architecture and Verification "I would never allow AI to invent architecture. Writing AI agentic coding, Vibecoding, whatever coding—without proper verification and properly setting expectations of what you want to get as a result—that's the main mistake." Sergey identifies two non-negotiables. First, never let AI invent architecture. Use framework conventions (Rails, etc.) to constrain AI's choices. Leverage existing code generators and scaffolding. Provide explicit architectural guidelines in planning steps. Store iteration-specific instructions where AI can reference them. The framework becomes the guardrails that prevent AI from making structural decisions it's not equipped to make. Second, always verify AI output. Even if you don't want to look at code, you must validate that it meets requirements. This might be through tests, manual review, or automated checks—but skipping verification is the fundamental mistake. These two rules—human-defined architecture and mandatory verification—separate successful AI-assisted development from technical debt generation. Prototype vs. Production: Two Different Workflows "When you pair as an architect or a really senior engineer who can implement it by himself, but just wants to save time, you do the pair programming with AI, and the AI kind of ships a draft, and rapid prototype." Sergey distinguishes clearly between prototype and production development. For MVPs and rapid prototypes, a senior architect pairs with AI to create drafts quickly—this is where speed matters most. For production code, teams add more iterative testing and polishing after AI generates initial implementation. The key is being explicit about which mode you're in. The biggest anti-pattern is treating prototype code as production-ready without the necessary validation and hardening steps. When building production systems, Sergey applies the full transactional model: atomic commits, comprehensive tests, architectural constraints, and rollback strategies. For prototypes, speed takes priority, but the architectural knowledge still comes from humans, not AI. The Future: AI Literacy as Mandatory "Being a software engineer and trying to get a new job, it's gonna be a mandatory requirement for you to understand how to use AI for coding. So it's not enough to just be a good engineer." Sergey sees AI-assisted coding literacy becoming as fundamental as Git proficiency. Future engineering jobs will require demonstrating effective AI collaboration, not just traditional coding skills. We're reaching good performance levels with AI models—now the challenge is learning to use them efficiently. This means frameworks and standardized patterns for AI-assisted development will emerge and consolidate. Approaches like AAID, SpecKit, and others represent early attempts to create these patterns. Sergey expects architectural patterns for AI-assisted development to standardize, similar to how design patterns emerged in object-oriented programming. The human remains the bottleneck—for domain knowledge, business requirements, and architectural guidance—but the implementation mechanics shift heavily toward AI collaboration. Resources for Practitioners "We are reaching a good performance level of AI models, and now we need to guide it to make it impactful. It's a great tool, now we need to understand how to make it impactful." Sergey recommends Obie Fernandez's work on "Patterns of Application Development Using AI," particularly valuable for Ruby and Rails developers but applicable broadly. He references Andrey Karpathy's original vibecoding post and emphasizes Extreme Programming practices as foundational. The tools he uses—Cursor and Claude Code—support custom planning steps and context management. But more important than tools is the mindset: we have powerful AI capabilities now, and the focus must shift to efficient usage patterns. This means experimenting with workflows, documenting what works, and sharing patterns with the community. Sergey himself shares case studies on LinkedIn and travels extensively speaking about these approaches, contributing to the collective learning happening in real-time. About Sergey Sergyenko Sergey is the CEO of Cybergizer, a dynamic software development agency with offices in Vilnius, Lithuania. Specializing in MVPs with zero cash requirements, Cybergizer offers top-tier CTO services and startup teams. Their tech stack includes Ruby, Rails, Elixir, and ReactJS. Sergey was also a featured speaker at the Global Agile Summit, and you can find his talk available in your membership area. If you are not a member don't worry, you can get the 1-month trial and watch the whole conference. You can cancel at any time. You can link with Sergey Sergyenko on LinkedIn.
Sverige