From Code Writing to Agent Orchestration: Andrej Karpathy on Software's AI-Driven Future

During a fireside chat at Sequoia Ascent 2026, Andrej Karpathy, co-founder of OpenAI and leader of Eureka Labs, outlined a fundamental shift in how software is built. He argues that large language models are no longer tools that accelerate existing workflows—they are becoming a new operating layer for digital work, reorganizing how tasks are approached and executed. Karpathy observed a qualitative leap in AI agent capabilities around December 2025, when generated code chunks became significantly larger, more coherent, and more reliable. This shift has redefined what it means to be a programmer: the role is evolving from writing code to orchestrating agents. He frames this transition as "Software 3.0," where an LLM interprets and performs computation over digital context, adapting to environments with less precision but far greater flexibility than traditional programming. His MenuGen example illustrates this reorganization starkly. A traditional web app requires a full stack—frontend, APIs, image generation, deployment, and infrastructure. In Software 3.0, a multimodal model can take a menu photograph and directly generate dish images overlaid on it, eliminating most of the software scaffolding. This principle extends beyond coding: some applications should cease to exist as discrete software products. The implication for founders is profound—AI is not merely a faster way to build old applications; it is rendering some of them unnecessary. Karpathy emphasizes that model capability depends not only on task verifiability but also on how heavily the task was emphasized during training and reinforcement learning. Coding agents excel because they receive immediate feedback—tests pass or fail—while other domains may collapse unexpectedly. For founders, the strategic question is whether their domain sits within the model's trained rails. If not, they may need enhanced context, fine-tuning, or proprietary reinforcement learning environments. The contrast between "vibe coding" (suitable for prototypes) and "agentic engineering" (required for production systems) clarifies that the future is not programmer replacement but transformation. Agentic engineers design specifications, supervise plans, and create evaluation loops. Critically, human understanding remains irreplaceable—knowing what is worth building, recognizing suspicious results, and judging acceptable tradeoffs. This understanding is why Karpathy champions LLM knowledge bases as tools for transforming information into comprehension, not merely answering questions.