Vibe coding, what are we not talking about?

"Vibe coding" has become a polarizing concept in software engineering conversations. Advocates echo this as a path to flow, where you can use AI's buoyancy for turning intent to outcome. Detractors dismiss this as a path to unmaintainable systems and production disasters. But are these the right questions to ask?

When experienced engineers push back against vibe coding, their resistance comes from:

• AI-generated code feels "mostly right" but might contain subtle quirks that are difficult to detect
• AI systems might produce functional but inelegant code that accumulates technical debt over time
• Coding agents require significant tuning and setup to be effective in large, complex codebases
• Lack of production-specific context in AI-generated output, creates risk of breaking live systems

However, these concerns point to a much deeper challenge: the gap between what we can "vibe code" (individual components) and what we actually need to manage (production systems that encompass code, infrastructure, and operational complexity).

What is this “vibe coding” anyway?

Traditional programming requires dual-layer thinking: Your brain must simultaneously hold the high-level intent ("I want users to see their purchase history") AND the low-level implementation details ("I need to join the users table with the orders table, handle pagination, format timestamps..."). You're constantly switching between "what do I want this to do?" and "how do I make the machine understand this?".

Vibe coding collapses this thinking into a single layer focused on intent and system design. When you describe what you want to an AI, your brain can dedicate its full capacity to understand the problem domain, user needs, and architectural decisions. And note, even with advanced vibe coding, engineers remain crucial, as they guide and control the AI's output, using it primarily for high-level tasks and structured environments.

The core tenets of vibe coding are:

• Focus on the "what" and "why," while the AI handles the "how." The goal is to stay engaged with the problem and the product's purpose.
• Delegation of implementation details. You can conserve mental energy for higher-level problem-solving and creative thinking.
• Rapid feedback loops for fast iteration and prototyping. You can test ideas, gather feedback, and pivot without getting bogged down in syntax.

The catch? This requires you to trust the AI systems you are relying on. If you have to pause and ask, "But does the AI know about the weird edge case in the legacy billing module?" or "What are the downstream impacts of changing this API?" you’re right back in dual-layer thinking, breaking the “flow” state.

Does it feel riskier to vibe in your current production systems?

The data on vibe coding points to an interesting insight.

• We hear that "in the current batch of Y Combinator-funded startups, 25% indicated that more than 95% of their codebase is AI-generated."
• At the same time, giants like Google and Meta are adopting AI more cautiously. Sundar Pichai states, "More than a quarter of new code at Google is generated by AI," which is a massive figure but a far cry from the 95% seen in startups.

Why the difference?

• Greenfield and the benefit of zero baggage: A startup has a "greenfield" project. The codebase is small, dependencies are known, and there is no legacy system to break. The AI's context window can practically encompass the entire universe of the project.
• Brownfield and the weight of success: Google or Meta, on the other hand, operate on a "brownfield" of immense complexity. Their systems are layered with decades of code, hidden dependencies, and business logic that no single human, let alone an LLM, fully comprehends. The cost of failure isn't a buggy prototype; it's a global outage impacting brand, revenue, and user experience for billions globally. The AI doesn't just need to know the code; it needs to understand the organization's history, its past incidents, and its unwritten operational rules. It needs to understand the organizational context. This is the quicksand: the immense, implicit context of a successful, long-running system.

For an enterprise engineer, blindly trusting AI-generated code feels like negligence. Their job is straightforward conceptually but extremely hard in practice: build new things fast without breaking anything already in production. This defensive posture is a rational response to the environment they work in.

Hard truth: We are already working with production systems we cannot maintain

YC startups showing "95% AI-generated code" aren't just working with smaller systems. They're building systems where "coding" encompasses only a fraction of the total complexity. But "coding" in reality is also about how it uses your existing systems and operates together cohesively. That's the real hard problem. New code is just one component of a much larger system, and currently, it's the only thing we can effectively "vibe code."

Meanwhile, enterprises with "~25% AI-generated code" are trying to retrofit AI collaboration onto systems designed for human comprehension. It's like trying to drive a car on train tracks.

But here's the kicker: our enterprise systems have already grown beyond human cognitive capacity. An average enterprise system has millions of lines of code, thousands of dependencies, and decades of accumulated intuition. No single human understands it completely. We've been operating in this territory for years.

The final and most significant blocker is tribal knowledge or engineering intuition. This is the unwritten wisdom held by your most seasoned engineers. The intuitive understanding of which services are brittle, what alerts are secretly critical, and how the system really behaves under pressure. This knowledge is rarely documented, and if it is, it becomes outdated if not maintained daily. Also, this knowledge cannot be fed into any legacy tool or platform.

Resistance to vibe coding isn't just about production safety. It’s about confronting the uncomfortable truth that these systems are already too complex for pure human management. AI systems aren’t just adding to the production complexity. In fact, they are revealing that maintainability was already an illusion. It's also about understanding that, at least for the foreseeable future, AI isn’t capable of replacing human programmers for complex projects, but what are the high-level tasks or operations within structured environments that it can do now with greater speed and efficiency?

A counterpoint: Should you be able to understand your production? Or just navigate it well?

The startup vs. enterprise adoption difference isn't really about codebase size. It’s about our approach to systems. When seasoned engineers say AI-generated code is unmaintainable, in some ways, we are also assuming that humans are always going to be responsible for reading and maintaining production.

Consider this thought: an AI system writes code that implements your requirements but uses unconventional patterns. Is it "unmaintainable" only because you assume humans need to modify it manually? But if you have another AI system that can read, understand, and modify your previous work, then “vibing-maintaining” becomes an AI-to-AI communication problem.

We're moving from human-readable code to AI-readable systems based on human design and guardrails. The question shouldn’t be "can I understand this code?" but "can I use AI to understand this system's behavior well enough to direct its evolution?" This shift in thinking is not trivial. It's about confronting the reality that most human changes are also context-blind. We just pretend otherwise because admitting the full scope of what we don’t know about our production systems is too uncomfortable.

The real maintainability question now becomes: can I effectively collaborate with AI to evolve this system over time? This requires entirely different skills from operating production.

Using Agentic AI for both coding efficiencies and production reliability

It's time we accept that complete human understanding of production systems is impossible, and we need to rely on AI systems that are our digital teammates that can navigate the complexity autonomously and at scale. This represents a philosophical shift from "systems should be understandable" to "systems should be navigable". Agentic AI systems are designed not just to write code, but to understand the production systems. By integrating directly with the operational data (traces, logs, metrics, and incident history), it builds a complete model of how your system actually works, including its hidden dependencies and learned behaviors. It doesn't need a human to explain the tribal knowledge because it learns it directly and continuously from every interaction. When Agentic AI systems work with you in production, it's not a generic solution from the internet; it's a reasoned recommendation grounded in the reality of your system. It can investigate incidents autonomously, explain complex dependencies, and guide resolutions with an awareness that, until now, only existed in the minds of your senior staff. This frees the engineer from the fear of the unknown, finally creating the psychological control required to "give in to the vibe."

About Resolve AI

Resolve AI is the agentic AI company for software engineering founded by the co-creators of OpenTelemetry.

Resolve AI understands your production environments, reasons like your seasoned engineers, and learns from every interaction to give your teams decisive control over on-call incidents with autonomous investigations and clear resolution guidance. Resolve AI also helps you ship quality code faster and improve reliability by revealing hidden system context and operational behaviors. With Resolve AI, customers like DataStax, Tubi, and Rappi, have increased engineering velocity and systems reliability by putting machines on-call for humans and letting engineers just code.