How Cognition Built the World's First AI Coding Agent—Before Claude Code
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this compelling conversation, Dan Shipper sits down with Scott Wu, the co-founder and CEO of Cognition Labs (makers of Devon AI), to explore how AI is fundamentally reshaping software engineering. (00:40) Scott shares his provocative stance that AGI is already here—or at least much closer than most people think—and discusses how Devon, which has reached $73 million in ARR, is changing what it means to be a software engineer in the AI era. The conversation delves deep into the evolving landscape of AI coding tools, from IDE assistants to fully autonomous agents, and examines how programming will transform over the next decade.
- The episode explores the radical transformation of software engineering through AI, the different paradigms of AI coding tools, and Scott's unique perspective on what programming will look like when engineers can operate at higher levels of abstraction.
Speakers
Scott Wu
Scott Wu is the co-founder and CEO of Cognition Labs, the company behind Devon, an AI software engineer that has reached $73 million in ARR. Scott previously founded and ran Lunch Club for five years, giving him substantial startup experience before diving into the AI coding space. His company recently acquired Windsurf, a cursor competitor, positioning Cognition as one of the foremost players shaping the future of programming over the next 10-20 years.
Dan Shipper
Dan Shipper is the host of the AI and I podcast and a respected voice in the AI space. He brings deep technical understanding and practical experience to conversations about AI's impact on work and society. Dan focuses on extracting actionable insights from AI leaders and experts to help professionals understand and adapt to the rapidly changing landscape of artificial intelligence.
Key Takeaways
AI Coding Requires a Fundamentally Different Skill Set
The shift to AI-assisted programming isn't just about using new tools—it's about developing entirely new capabilities focused on higher-level thinking. (18:18) Scott explains that future engineers will need to excel at "really deeply understanding logical fundamentals, being able to break down problems and articulate the answers to them, being able to think about different strategic trade-offs, thinking about architectures." This represents a move from knowing obscure libraries and debugging Kubernetes to operating as technical architects who can direct AI agents effectively.
The Form Factor Revolution: From IDE to Agent Orchestration
The recent explosion of CLI-based AI tools like Claude Code represents a fundamental shift in how we interact with programming tools. (30:14) Scott notes that "the correct form factor is a pretty tight function of the capabilities" of the underlying AI models. As capabilities improve, the most effective interface isn't about augmenting your existing workflow—it's about "handing the reins over to your AI buddy to take the wheel of your computer," representing a full-send approach to agentic engineering.
The Sync-to-Async Workflow Will Define the Next Three Years
Professional software engineers will increasingly need to master the art of transitioning between synchronous collaboration with AI and asynchronous delegation to autonomous agents. (24:45) Scott predicts this hybrid phase will last about three years, where engineers use both IDEs for hands-on work and background agents for tasks that can run independently. The key skill becomes knowing "how you split up your tasks into which ones should go into which buckets" and seamlessly moving between these different modes of work.
Post-Training Beats Base Model Development for Specialized Applications
Rather than building models from scratch, the winning strategy for AI coding companies is intensive post-training on domain-specific tasks. (38:49) Scott explains that Cognition focuses on "teaching models things like predicting their own confidence, understanding specific engineering flows like pulling up Datadog logs to debug issues, and handling real-world software engineering practicalities." This approach allows startups to compete with tech giants by developing deep expertise in their specific domain rather than trying to match general-purpose model capabilities.
RL Environments Are the New Competitive Moat
The companies that will win in AI coding are those building the most sophisticated reinforcement learning environments that mirror real-world software engineering challenges. (41:57) Scott describes creating evaluations like setting up broken Grafana dashboards where "if you did not get the Grafana dashboard running, you will never be able to answer that question. And if you did get it running, you will always get the right answer." This tight feedback loop enables much more effective model training than generic approaches, making environment design a crucial competitive advantage.
Statistics & Facts
- Devon has reached $73 million in Annual Recurring Revenue (ARR) just a couple of years after its launch, demonstrating the rapid market adoption of AI software engineering tools. (00:29)
- Scott Wu ran his previous company, Lunch Club, for about five years before starting Cognition, giving him substantial entrepreneurial experience prior to entering the AI coding space. (09:04)
- Scott predicts that the hybrid phase of using both IDEs and autonomous agents will last approximately three years, after which programming will transition to primarily dictating requirements to AI agents. (24:48)
