Subscribe

In February 2025, Andrej Karpathy, co-founder of OpenAI, gave a name to something that was already happening everywhere: vibe coding. Suddenly, anyone with a laptop, an afternoon, and a $20 monthly subscription could prompt a large language model and have an app running within days. Computer science graduates experienced a wave of career anxiety, and businesses began asking what the org chart might look like with fewer engineers. At the center of these responses sits a simpler, more lasting question: What do we want AI collaboration in the workplace to look like?

Vibe coding was, for a moment, an answer. It promised speed, flexibility, and automation, despite being contingent on a high degree of trust in the AI models. In the year since, however, it’s only revealed what happens to software when you remove the engineering part.

The engineering pipeline for any new product or feature is robust for a reason. Security review, unit testing, and staged deployment are just as important for software as the code itself. When vibe coding collapses that pipeline for the sake of speed, the quality of the product suffers.

For some, this is a reason to scrap vibe coding altogether. But Lakshya Agarwal, a McGill University alumnus and forward-deployed engineer at a prominent AI startup called Tavily, has a different take. A vibe coder himself, he sees the technology’s potential benefits – and its risks.

“The capability of these models is unparalleled today,” said Agarwal in an interview on the McGill Delve podcast.

They can accelerate iteration and compress the time needed to produce a prototype, he explained. And to someone who doesn’t code at all, this can feel like magic.

But vibe coding also comes with risk. Models can easily make mistakes, said Agarwal. If users trust the model too much, they can end up with a subpar product.

So the problem with vibe coding, Agarwal emphasized, isn’t the AI agents. It’s the users. They can easily ask for too much from AI models with too little oversight.

For him, to blindly trust AI models would be to abandon some of the core principles of software engineering, such as testing, iterating, and accountability for the final product.

As a result, vibe coders increasingly recognize the need for more structure around their AI agent workflow. Enter: agentic engineering, a term coined by Karpathy only one year after the introduction of vibe coding.

If vibe coding forgets the core tenets of software engineering, agentic engineering reimagines them. It encompasses the same software engineering workflows and principles but now working side by side with AI at each step. This is essential for organizations hoping to maximize the benefits of vibe coding while avoiding its worst pitfalls.

Don’t trust the vibes

For organizations to collaborate with AI safely, Agarwal said, they must have robust human oversight over the models. This is the heart of agentic engineering.

“You, as the human or the user, are the one who is deciding what is correct and what’s wrong,” he said.

The engineer prompts the agent, but also validates and approves what it produces. They have to understand what the agent built well enough to take responsibility for it. That balance, he believes, is how organizations can maximize output while minimizing risk.

For Agarwal, agentic engineering is also an opportunity to change our mentality around AI collaboration.

“It isn’t about compressing speed,” he said. “It’s about solving harder problems with more help.”

In other words, the speed gained when using AI is significant, but it matters more what is done with the extra time. When software engineers can delegate the actual line-by-line writing of code to an AI agent, they’re freed up to focus on what to build and why, explained Agarwal. They’re engineers first and foremost. A civil engineer doesn’t spend their days mixing concrete, they spend them deciding where, how and why to pour it. Code is similar: a medium rather than the work itself. The time agentic engineering frees up doesn’t go towards writing more code; it goes to tackling harder problems, asking better questions, and finding more creative solutions.

Invest in culture

Your organization’s ability to benefit from AI will depend on your organization’s culture, according to Agarwal. Do people on your team feel safe to experiment, to push boundaries, to take ownership of a feature they want to see built? Without psychological safety, people doubt their ideas before they’ve even voiced them, and innovation benefits from the breadth of everyone’s suggestions.

This has always mattered. Now, however, with AI, the only thing standing between an idea and a new prototype is whether someone feels confident enough to try.

Agarwal acknowledged, however, that company culture isn’t a failsafe buffer in the face of disruption. Some roles will not survive the transition. Others will inevitably change shape.

This, he said, is nothing new.

“Throughout the history of humanity, at least since we’ve had the concept of economy, jobs have come and gone.”

His own role at Tavily reflects this. He spends more time with clients and sales work than a software engineer would have just a few years ago.

But AI collaboration shouldn’t mean wholesale replacement of a software engineering team either, argued Agarwal. The people most equipped to navigate this transition are those already inside the organization. They know the processes, edge cases, and customers.

The best thing an organization can do, he said, is invest in training — teaching people to collaborate with AI, to trust it where it earns trust, and to interrogate it where it doesn’t. The organizations that will benefit are the ones supporting their employees and building a culture in which AI expands what people can do rather than reducing how many people are needed to do it.