Vibe Coding and the Rise of AI Agents: The Future of Software Development is Here

Ben Lorica

Steve Yegge on Vibe Coding, Agent Programming, AI Engineering, and the Future of Software Development.

Subscribe: Apple • Spotify • Overcast • Pocket Casts • AntennaPod • Podcast Addict • Amazon •  RSS.

Steve Yegge is an evangelist at Sourcegraph, a startups that is industrializing software development with AI agents. This episode explores the paradigm shift in software development with the rise of “vibe coding” and AI agents, moving beyond traditional code completion. It discusses how developers are transitioning from line-by-line coding to orchestrating AI, emphasizing the crucial need for trust, verification, and new skill sets like AI engineering and humanities. The conversation highlights the transformative potential of coding agents to dramatically increase productivity and reshape the future of software development, while also addressing challenges and offering practical advice for developers to adapt and thrive.  [This episode originally aired on Generative AI in the Real World, a podcast series I’m hosting for O’Reilly.]

Subscribe to the Gradient Flow Newsletter

Interview highlights – key sections from the video version:

Related content:

Support our work by subscribing to our newsletter📩

Transcript

Below is a heavily edited excerpt, in Question & Answer format.

What is chat-oriented programming (CHOP) or “vibe coding” and how does it differ from traditional code completions?

Chat-oriented programming, which I coined about a year ago, has been repackaged as “vibe coding” by Andre Karpathy from OpenAI. Vibe coding is fundamentally about having AI do your work for you, but it doesn’t mean turning your brain off completely. The level of attention you pay depends entirely on the problem you’re working on.

If you’re doing weekend prototyping or exploring multiple options, you can be less rigorous. But when hardening code for production, you continue using AI while demanding excellence from it. This represents a significant evolution from code completions, which are now considered outdated. If you’re still measuring completion acceptance rates, you’re behind the curve. Chat programming is current, but agent programming is already rocketing past that approach with exponentially better results.

How does vibe coding change the role of a traditional software developer?

The traditional software developer’s role is transforming from writing code line-by-line to becoming a high-level orchestrator or strategic guide. This is one of the most challenging transitions the industry has ever faced. You’re no longer writing code yourself, but you remain fully accountable for it. You can’t blame AI for production issues – that’s like blaming your compiler for an outage.

This shift is especially difficult when using coding agents like Claude Code, where you might have multiple agents simultaneously working on different tasks – one handling bug backlog grooming, another coding a new feature, and another doing code reviews. You become massively more productive, but you’re not a programmer in the traditional sense anymore, which is psychologically challenging for many developers who built their identity around coding skills.

What is the trust dimension in AI-assisted coding and how should developers approach it?

Trust is the number one factor holding back industry adoption. Many senior engineers who haven’t embraced AI don’t try hard enough with these tools. They’ll give the AI their hardest interview question, and when it fails, they dismiss it entirely. That’s like giving an intern your toughest problem and firing them immediately if they can’t solve it.

The reality is that LLMs don’t work perfectly the first time. If they have a 20% hallucination rate, it doesn’t mean one out of five answers is wrong – it means portions of answers contain errors of varying severity. You need to validate everything, but that doesn’t make the tools less valuable. Software development with AI becomes a verification process rather than a creation process, and this mental shift is difficult but necessary.

What are the core skills developers need to embrace in the world of vibe coding and AI-assisted development?

First, developers need to understand AI engineering thoroughly. I recommend the O’Reilly AI Engineering book by Chip Huyen – all engineers should know its contents inside out. You can make room in your brain by deprioritizing programming language syntax, as LLMs will increasingly handle code writing.

Second, learn vibe coding properly instead of resisting it. It’s like learning to ride a motorcycle – you might fall off a few times, but once you master it, the productivity gains are extraordinary.

Third, prompt engineering, which differs from chat-oriented programming. Prompt engineering focuses on static prompts meant to last a long time, while chat programming involves creating dynamic, throwaway prompts to nudge the AI in the right direction.

Finally, humanities skills are becoming essential. As we increasingly manage AI agents, communication, coordination, logical thinking, and task breakdown – skills typically associated with humanities rather than engineering degrees – become crucial.

How do coding agents differ from chat-based programming, and what advantages do they offer?

Coding agents represent a quantum leap beyond chat programming. If chat programming is 5x more powerful than manual coding, agent coding is easily 5x faster than chat programming. The key difference is that agents operate in a persistent loop – they’re designed to brute-force solutions by burning through computational tokens.

For example, when you assign a JIRA ticket to Claude Code, it will first figure out how to access the ticket, whether through web browsing, command-line tools, or REST protocols. It’ll examine your codebase using familiar tools like grep and ls, develop an execution plan, implement the solution, write tests, update documentation, and even offer to commit the changes. You can interrupt and redirect at any point, but the agent handles the entire workflow that you previously had to manually direct in chat programming.

These agents are already as productive as human developers at a fraction of the cost – about $10/hour. The ability to run multiple agents simultaneously will dramatically increase productivity, making their adoption a competitive necessity.

What challenges do LLMs face when handling context across large software projects?

Right now, LLMs struggle with large project context without Model Context Protocol (MCP) servers, which can be described as “the new HTTP.” MCP is how intellectual property assets at companies will get wired together with AI, allowing AI systems to help with them.

Without MCP, LLMs have limited visibility into your broader system. They can perhaps read your Git repo, but what about databases, logs, documentation systems, and other components? This is fundamentally a Retrieval-Augmented Generation (RAG) problem – AI needs access to your data to help you effectively, and currently, developers must manually provide this context, which is inefficient.

Companies need to start writing their own MCP servers to connect their intellectual assets with AI systems. This infrastructure will be crucial for enabling more sophisticated AI assistance across larger software projects.

How do different LLMs compare for coding tasks, and should developers use multiple models?

Different models have distinct strengths, and using multiple models can be valuable for verification. Getting a “second opinion” from another model is similar to consulting another doctor. This approach helps validate solutions, especially when working in unfamiliar domains or programming languages.

Currently, Claude appears to be the best at coding, even though the benchmarks don’t always reflect this. This suggests there are intangible aspects to model performance that aren’t captured by standard metrics. However, the models are rapidly converging in capability, and with coding agents that brute-force solutions through multiple attempts, the specific model may become less important over time.

The industry is heading toward commoditization of the underlying models, but we’re not there yet. For now, having access to multiple models provides pragmatic advantages for verification and different problem domains.

What is the projected state of coding agents in the next 6-12 months, and what are the financial implications?

We’re seeing six overlapping waves of AI coding modalities, with traditional programming declining and expected to be largely obsolete by 2027. Completions provided a small boost through 2023, chat programming dominated 2024 with about a 5x productivity increase, and now agent programming is taking over with another 5x improvement.

In the next 6-9 months, we’ll see more sophisticated tools for managing multiple coding agents simultaneously. Currently, each agent is roughly as productive as a human developer at about $10/hour. With improved tools, a developer might run 5 agents at once for $50/hour, becoming 3-5x more productive.

This creates a significant financial challenge for companies that have already completed their fiscal planning for 2025. They suddenly need to allocate substantial budget for token spend that wasn’t anticipated. Companies with deep pockets can simply pay to play, but others will face difficult decisions – either absorb the costs, fall behind competitors, or reduce headcount to cover the new expenses. Some companies have already laid off 30% of their engineers who wouldn’t adopt AI.

How will AI-assisted coding affect computer science education and career progression?

Computer science education will need to evolve, but the fundamentals remain valuable. When assembly language was replaced by higher-level languages, people worried programming skills would deteriorate, but instead, the field expanded and jobs increased. The same transformation is happening now – we’re moving up the abstraction ladder.

CS degrees will likely adjust by making programming language courses optional while adding AI engineering courses. The core focus will shift toward architecture, algorithms, and system design – areas where understanding remains crucial even with AI assistance.

Career progression is also changing. Interestingly, many senior developers are struggling more with this transition than juniors. Junior developers, faced with competitive job markets, are embracing AI as mentors and rapidly adapting to the new paradigm. They’re using AI to accelerate their learning and transform themselves into the “new engineer” – someone who can effectively collaborate with and direct AI systems.

What practical steps should developers take to stay relevant in the age of AI-assisted coding?

  1. Learn AI engineering thoroughly – get the O’Reilly AI Engineering book and master its contents
  2. Embrace vibe coding instead of resisting it – the productivity gains are too significant to ignore
  3. Start working with coding agents immediately – try Claude Code or similar tools even though they’re still in early stages
  4. Practice with simpler tools first if needed – use Claude for coding problems to get comfortable with the AI collaboration process
  5. Develop humanities skills alongside technical ones – communication, coordination, and task breakdown become increasingly important
  6. Don’t just dabble – commit to making a serious transition, as half-measures won’t be sufficient
  7. Understand that AI won’t replace software development, but developers who use AI will replace those who don’t
  8. Use multiple models for verification and different problem domains
  9. Start learning about MCP and building infrastructure to connect your systems with AI
  10. Be prepared for continued exponential change – what seems cutting-edge today may be obsolete in months

The pace of change is unprecedented and exponential. Organizations and individuals that recognize this and adapt quickly will thrive, while those in denial face increasingly difficult prospects in the software development industry.