From Code Generation to AI-Native: Insights from the Front Lines
Just yesterday, I analysed the AWS guide on “vibe coding,” noting how it formalises AI-assisted development for the enterprise. It represents a shift from individual craft to scalable, process-driven systems. A recent panel discussion with experts from Anthropic, Trae, and Replit provides a complementary view, offering a look at the best practices emerging from the very frontier of this technology.
This discussion moves the conversation from AI-assisted coding to what can be called “AI-native” development. It reinforces the idea that the developer’s role is fundamentally changing, and it provides practical strategies that bridge the gap between solo experimentation and enterprise-grade systems.
The Core Shift: AI as a Collaborator
The panellists—Paxton Maeder-York from Anthropic, Chao Peng from Trae, and Michele Catasta from Replit—unanimously framed AI not as a tool, but as a collaborative partner. This is more than a semantic distinction; it underpins a new development paradigm.
Paxton Maeder-York (Anthropic): “I typically say I collaborated with Claude… Claude and I are best friends and we do a lot of work together. We’re very much encouraged at Anthropic to try to automate our entire job away with Claude.”
This perspective aligns with the principles I have explored previously: the human acts as the architect, directing a powerful collaborator. The panel’s insights show how this collaboration works in practice.
Battle-Tested Best Practices
The discussion yielded several concrete strategies that are being used right now.
Strategic Context Management: Even with large context windows, managing what the AI knows is critical. Paxton shared a practical tip: strategically “compacting” a conversation’s history into a summary file (like
claude.md) to maintain essential context. This is a more dynamic, developer-led version of the formal requirement documents proposed in the AWS guide. He also offered a simple but effective prompt suffix: “don’t break anything and don’t change anything else.”Multi-Agent Orchestration: The concept of using specialised sub-agents, which I have noted from developers like Zach Wills, is now a key strategy inside Anthropic. Paxton confirmed that their internal experts run multiple AI agents in parallel, each working on a different part of the codebase. This is the practical application of turning the developer into an orchestrator of an AI team.
The Resurgence of Test-Driven Development (TDD): The panel offered a powerful solution to the problem of AI-generated code quality. Both the AWS guide and my own experience suggest that AI assistants are poor at designing meaningful tests. Chao Peng argued for a TDD revival: the human defines the tests, and the AI’s job is to write code that passes them. This creates a clear feedback loop and makes debugging far more efficient for both the human and the AI.
The Final 20%: Enterprise Readiness
The panel was realistic about the challenges of deploying AI-built applications. Michele Catasta pointed out the “foot guns” that coding agents expose and the complexities beyond pure code generation.
Michele Catasta (Replit): “There are many other things that are important: how do you deploy? Data governance, security and so forth?”
This directly explains why the enterprise-focused approach seen in the AWS guide is so process-heavy. The formalisms and security guardrails are not bureaucracy; they are the necessary systems to manage the final, difficult 20% of getting a product into production reliably and securely.
The Engineer’s Evolving Role: From Coder to Specifier
Ultimately, the panel confirmed a conclusion I have been moving toward: the core work of a senior engineer is shifting from implementation to specification.
Michele Catasta (Replit): “I think it’s already the case. Like the more you grow as an engineer in your career, the more you write specs rather than code. That’s where the important insights are.”
This is the end-game of the “human as architect” model. Our primary value is no longer in writing lines of code, but in defining the problem with such precision that an AI collaborator can execute the solution. The innovator’s agile mindset and the enterprise’s formal processes are both methods for achieving this clarity of specification. The most effective developers will be those who master this skill of architectural definition, regardless of the tools they use.
