Why Karpathy Says Vibe Coding Isn’t Dead and Software 3.0 Is Giving Rise to Agentic Engineering

1. Why Karpathy’s View Matters Now

Karpathy first named Vibe Coding in February 2025, then articulated Software 3.0 at Sequoia AI Ascent on 20 April 2026, and a month later announced his move to Anthropic (19 May 2026). His credibility stems from early OpenAI research, leading Tesla’s AI work, and ongoing public talks on LLM‑based programming.

The core thread across these milestones is a single question: as model capability rises, how should software work be layered?

2. Software 3.0 Shifts the Programming Target

Software 1.0 encodes explicit rules; Software 2.0 embeds behavior in trained weights. Software 3.0 moves a step further: developers program LLMs with prompts, context, examples, tools, and memory. Karpathy treats prompts as the new program and emphasizes the context window as a key lever.

Consequently, the product must decide what to show the model, which APIs it may call, how to collect feedback, and when to hand control back to a human.

MenuGen illustrates the shift: the 2025 version was a full‑stack web app generated by Cursor and Claude; the 2026 demo fed a menu image to a multimodal model, letting the model draw the dishes directly on the picture, bypassing OCR, page‑interaction, and intermediate services.

Software 3.0 does not discard earlier paradigms—identity, payment, audit remain in the deterministic Software 1.0 layer, the underlying model stays in Software 2.0, while natural‑language interaction and agent orchestration belong to 3.0.

3. Vibe Coding Is Not Dead, It Has Been Split

Vibe Coding, described in February 2025, is a low‑overhead workflow: voice commands, no diff reading, agents make changes, errors are pasted back, and the code itself is often forgotten. Its goal is to turn a fleeting idea into a runnable prototype quickly, without guaranteeing long‑term reliability.

This lowers the entry barrier, allowing non‑engineers to build demos and enabling professional developers to experiment cheaply. However, when a product moves beyond exploration, the full spec‑test‑review pipeline becomes essential.

Karpathy’s 2026 conclusion was modest: for serious web apps, staying in pure Vibe mode is unsafe. He therefore coined “Agentic Engineering” to describe a professional workflow where the agent generates code but humans retain specification, orchestration, supervision, review, and final responsibility.

4. Why Pure Vibe Is Insufficient

Karpathy argues that model ability is “jagged”: strong on large codebases or bug‑fixing tasks but prone to basic commonsense errors (e.g., suggesting a user walk 50 m to a car‑wash). This “Ghost vs. Animal” view means performance on one task does not guarantee reliability on adjacent tasks.

In low‑risk scenarios, Vibe Coding can tolerate such jaggedness because the output can be discarded. In high‑risk contexts—payments, authentication, data privacy—Agentic Engineering’s added verification, human oversight, and responsibility become mandatory.

5. Verifiability Is a Tool, Not the Core Thesis

Agentic Engineering relies on tests, reviews, and feedback loops, but these are only instruments. Verifiability helps decide which tasks can grant agents higher autonomy: compile‑able code, runnable tests, resettable environments, and comparable results. Tasks that lack clear evaluation, have irreversible failure, or involve business‑critical decisions require tighter human control.

6. Human Skill Must Shift From Prompting to Understanding

As code generation becomes cheap, the valuable human skill moves from writing prompts to grasping underlying concepts—storage semantics, permission models, failure recovery, and business logic. Karpathy quotes a tweet: “You can outsource your thinking, but you can’t outsource your understanding,” emphasizing that agents can handle execution, but humans must retain deep comprehension.

7. What Product Managers Should Change

First, verify whether the product itself is still needed; ask if the user needs a full workflow or just an input‑to‑output transformation.

Separate demo viability from production readiness; a prototype’s speed does not imply compliance with identity, payment, or security requirements.

Configure agent autonomy based on risk: reversible, auto‑verifiable tasks may receive longer chains; high‑risk tasks (finance, privacy) need shortened steps and manual sign‑offs.

Define the agent’s contract: input spec, orchestration, permission boundaries, testing feedback, review, and final accountability.

8. The Overall Takeaway

Software 3.0 makes natural language, context, and tool calls the new programming material. Vibe Coding leverages this to accelerate prototyping; Agentic Engineering re‑integrates specification, validation, review, and responsibility for production‑grade work. The split reflects differing failure‑cost profiles, not the death of Vibe Coding. Karpathy’s move to Anthropic places him back in frontier R&D, but the concrete conclusion for practitioners is clear: use Vibe Coding for rapid prototypes, switch to Agentic Engineering when real users bear risk.