Last November, we wrote about why the vibes are off. AI coding tools had exploded in popularity, everyone was building apps from prompts, and the enterprise software world was trying to figure out whether to embrace it or ignore it.
Our take: vibe coding is real, and the productivity gains are real. But the code that comes out the other end still needs experienced humans to take responsibility for it.
Since then, the conversation has moved faster than any of us could have imagined.
In February 2026, Andrej Karpathy—the AI researcher who coined 'vibe coding' in the first place —effectively retired the term. On the one-year anniversary of the phrase, he wrote that it had been coined casually, and that it was appropriate for a period when the tools were far less capable. Today, he says, something different is happening. He now prefers the term agentic engineering.
If you lead a team that builds software, or if you're evaluating external partners for a project, this shift matters more than a vocabulary change. Here's what it actually means.
What is agentic engineering, in plain terms?
Vibe coding, as originally conceived, was about prompting an AI to write code and accepting whatever came back. You described what you wanted, the model generated it, and you ran it. If it broke, you pasted the error back in. It was fast, accessible, and great for prototypes. It was not, as we argued, production-ready on its own.
Agentic engineering is something more structured. Rather than asking an AI to generate a block of code once, you're orchestrating AI agents. These are systems that can plan, execute, test, and iterate through a task with a degree of autonomy. The human developer isn't writing every line, but they're not just typing prompts either. They're acting as the architect and the reviewer.
They set the constraints, define the quality gates, and own the outcome. Just like a product manager.
"You are not writing the code directly 99% of the time," Karpathy wrote. "You are orchestrating agents who do and acting as oversight."
The word engineering in agentic engineering is doing a lot of work. It signals that this isn't just a productivity hack. It's a professional discipline with real skill requirements and real accountability.
Why this shift matters for enterprise software projects
If you've been cautious about AI-generated code, here's the relevant context: most of the concerns were legitimate, and they still apply.
Early vibe-coded systems tended toward the same set of failure modes. Security vulnerabilities built in from the start. Architectural decisions made by a model with no knowledge of your business logic. Code that works today and becomes unmaintainable in six months because nobody on the team fully understands what it does. A December 2025 CodeRabbit analysis of open-source repositories found that AI-co-authored code contained measurably more major issues than human-written code.
Agentic engineering addresses these problems, but only when it's done with the right guardrails. The difference between a well-run agentic engineering workflow and a problematic one comes down to a few things:
- How clearly is the agent's task defined before it starts?
- What quality gates exist before output is accepted?
- Who reviews the code?
- Do they have the depth to evaluate it properly?
This is where the senior expertise problem becomes acute. Agentic engineering disproportionately benefits experienced engineers and teams. If you know what good software architecture looks like, you can leverage AI agents as a genuine force multiplier. If you don't have that foundation in place, you're not accelerating development. You're accelerating technical debt.
Three risks enterprise leaders should be asking about
As agentic tools become more embedded in how software gets built, here are the questions that should be on your radar if you're a CTO, VP of Engineering, or a decision-maker evaluating a software project.
- Security at scale. An AI agent capable of writing and merging hundreds of pull requests per week with a small percentage of vulnerabilities still creates a significant number of new vulnerabilities every week. Agentic systems need security gates built into the workflow, not bolted on after the fact. Ask any partner you're evaluating how security is embedded in their agentic processes.
- Architectural intent. AI agents are good at generating code that satisfies a specification. They are not good at understanding the larger system it lives in— the history of decisions, the constraints of your existing infrastructure, the business logic that shapes what the right answer is. This is why the planning phase of agentic engineering matters so much. Vague direction produces vague code. The human expertise that defines the architecture upfront is what determines whether the agent's output is useful or a liability.
- Cognitive debt. In 2025, the dominant risk in AI-assisted development was technical debt—code that worked but was hard to maintain. In 2026, the emerging risk is being called cognitive debt: the accumulated cost of poorly managed AI interactions, context loss, and agent behaviour that nobody on the team fully understands. If a system is built by agents and nobody can reason clearly about how it works, that's a governance problem that compounds over time.
What this means for working with an outsourced development partner
Here's the honest version of what the agentic engineering shift means if you're evaluating external teams for a software project.
The productivity gains from AI tools are real, and any development partner worth working with is using them. But using agentic workflows without the right human layer can be a problem.
Are there senior engineers reviewing agent output with the depth to catch what an automated test won't? Is security integrated into the process or treated as a final checklist item? Does the team have a clear methodology for defining architecture before agents are unleashed, or are they prompting and hoping?
At BitBakery, we've been working through this as a team in real time by experimenting with agentic tools, stress-testing outputs, and developing our own guardrails. The three principles we described in November still hold: AI augments, not replaces. Humans own every line of shipped code. Every use of AI must be driven by clarity, curiosity, and critical thinking.
Agentic engineering doesn't change those principles. It makes them more important.
The bottom line
The industry is moving fast and the terminology is moving with it. Vibe coding named a moment. Agentic engineering describes where professional software development is actually headed with AI doing more of the implementation work, humans providing the architecture, the oversight, and the accountability.
For enterprise organizations, the practical question is the same one it's always been: who is responsible for what ships? Agentic workflows don't change that question. They raise the stakes on it.
If you're thinking about how AI fits into your next software project—whether you're building from scratch, extending an existing system, or evaluating external partners—we'd be glad to talk through what responsible agentic development actually looks like in practice.
