From Vibe Coding to Agentic Engineering: Why Human Judgment Becomes the Real AI Advantage

Author’s Note and Disclaimer:

Zion Zhao Real Estate | 88844623 | 狮家社小赵 | wa.me/6588844623 |  https://linktr.ee/zionzhao

This post is for general information, education, and market literacy only. It does not constitute financial, investment, trading, legal, tax, accounting, or other professional advice, and is not an offer, solicitation, recommendation, or endorsement. Views expressed are personal, general in nature, and subject to change without notice. While reasonable care is taken, no representation or warranty is given as to accuracy, completeness, or reliability. Readers should conduct independent due diligence and seek professional advice. To the fullest extent permitted by law, no liability is accepted for any loss arising from reliance on this material. 

AI Is Not Replacing Software Talent. It Is Raising the Bar for Judgment, Verification and Responsibility

Andrej Karpathy’s message is not that coding is dead. It is that software work is being reorganised around a more powerful and more demanding discipline: agentic engineering. “Vibe coding” captured the first cultural wave of AI-assisted software creation, where users could describe what they wanted and let large language models generate the code. It raised the floor by allowing founders, students, designers, operators and non-technical builders to create working prototypes faster than ever before. Agentic engineering is the next and more serious phase. It raises the ceiling by asking professionals to direct AI agents while preserving quality, security, maintainability, architecture, privacy, accountability and judgment.

Karpathy’s Software 3.0 framework explains why this shift matters. Software 1.0 was explicit code written by humans. Software 2.0 was neural networks trained through data, objectives and weights. Software 3.0 is context, prompts, documents, tools, memory and workflows used to steer LLMs as programmable interpreters of digital information. The prompt is no longer just an instruction. It becomes a new kind of interface, a control layer and a form of programming. That is why AI is not merely accelerating existing software development. It is changing what software is.

His MenuGen example is particularly revealing. In the old paradigm, a menu-visualisation app requires OCR, image generation, interface design, deployment and multiple service integrations. In the new paradigm, a multimodal model may increasingly take a photo of a menu and directly transform it into an annotated visual output. The application layer can shrink, collapse or disappear. The strategic question is no longer simply “How do I build this app?” It is “Should this even be an app, or should it be a model-native workflow, agent task, tool call, or context-driven transformation?”

The key business insight is verifiability. AI advances fastest where outputs can be checked, tested, scored or rewarded. This explains why coding, mathematics, structured workflow automation, reconciliation and repeatable operational tasks are moving so quickly. Code can be compiled. Tests can be run. Outputs can be compared. Errors can be surfaced. Feedback can be turned into improvement loops. Research on AI-assisted software development and workplace productivity supports this direction, showing meaningful productivity gains while also highlighting uneven outcomes depending on task design, user expertise and organisational controls (Brynjolfsson et al., 2025; Peng et al., 2023). AI is leverage, not magic.

This is where Karpathy’s distinction becomes commercially important. Vibe coding may be enough for prototypes, experiments and internal tools. It is not enough for production-grade systems. A serious business cannot excuse weak security, poor architecture, broken identity logic, payment mistakes, privacy breaches or regulatory non-compliance by saying the AI wrote the code. The human team remains accountable. Agentic engineering therefore requires specification discipline, acceptance criteria, test suites, sandboxing, permission control, audit trails, code review, deployment safeguards and human approval for high-risk actions.

Karpathy’s “ghosts, not animals” metaphor adds another layer of realism. LLMs are not human-like minds with stable understanding, intrinsic motivation or grounded common sense. They are statistical, jagged, context-sensitive systems shaped by training data, reinforcement signals and tool access. They can be astonishingly strong in one domain and strangely weak in another. A model may refactor a complex codebase yet fail a basic practical reasoning question. This does not make AI useless. It means AI must be mapped, constrained and verified. The professional task is to know where the model is reliable, where it is brittle, where it needs tools, where it needs retrieval, and where human judgment must remain firmly in control.

That is why governance is not a side issue. It is the operating system of agentic AI. As agents receive more permissions, access more data and take more actions, organisations need clear rules on what they may do, what they may not do, when humans must approve, how outputs are logged, how failures are reviewed and how sensitive information is protected. Frameworks from NIST, OECD and OWASP all point toward the same conclusion: trustworthy AI requires validity, reliability, safety, security, accountability, privacy, transparency and human oversight (NIST, 2023; OECD, 2019; OWASP, 2025).

The professional lesson is sharp. Intelligence is getting cheaper, but understanding is getting more valuable. The winners of the AI era will not be those who merely prompt faster. They will be those who know what should be built, why it matters, what trade-offs are acceptable, how to specify the work, how to verify the output and when not to trust the machine. In that sense, the future does not belong to passive users of AI. It belongs to directors of intelligence.

For founders, executives, investors and professionals, Karpathy’s warning is clear: do not mistake automation for strategy. The opportunity is not just to add AI on top of old workflows. The opportunity is to redesign workflows around verifiable intelligence, agent-native infrastructure and human accountability. In the old world, leverage belonged to those who could write the code. In the agentic world, leverage belongs to those who can define the mission, constrain the agent, verify the result and understand the consequences.

References

Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics.

Karpathy, A. (2026). Andrej Karpathy: From vibe coding to agentic engineering. Sequoia AI Ascent 2026.

National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework: AI RMF 1.0.

OECD. (2019). OECD AI Principles.

OWASP. (2025). OWASP Top 10 for Large Language Model Applications.

Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot.

The End of Vibe Coding, the Rise of Agentic Engineering and the New Rules of Professional Work

Karpathy’s warning is clear: AI is not ending software, it is elevating it. Vibe coding expands access, while agentic engineering demands judgment, verification, security and accountability. As intelligence gets cheaper, understanding becomes the premium skill. The future belongs to professionals who can direct machines without surrendering responsibility.

Karpathy’s essay on “vibe coding” and agentic engineering is not only about technology. It is about the next phase of decision-making. In Singapore property, the same principle applies: tools are becoming faster, data is becoming cheaper, but judgment is becoming more valuable.

For buyers, AI can compare listings, analyse floor plans, estimate affordability and summarise market data. But it cannot fully understand your family needs, risk appetite, financing structure, school priorities, exit strategy or long-term wealth plan.

For sellers, technology can generate reports and marketing content. But pricing strategy, buyer psychology, negotiation timing and valuation defence still require experience, market feel and disciplined execution.

For landlords and tenants, automation can assist with documents and checks. But tenancy clauses, compliance, rent positioning, repair obligations and negotiation risks require professional judgment.

For investors, the lesson is even sharper. In a market shaped by interest rates, policy changes, supply pipelines, immigration flows, rental demand and global capital movements, information alone is not enough. You need someone who can interpret the signal from the noise.

This is where I bring value. I combine Singapore real estate practice with macroeconomics, asset allocation, portfolio thinking, legal awareness and market cycle analysis to help clients buy, sell, rent and invest with greater clarity.

If you are planning your next Singapore property move, engage me for a structured, data-informed and strategy-led consultation.

Like, collect and subscribe to my social media channels for more Singapore property insights, market education and practical decision frameworks.

Note: This content is for general education only and does not constitute financial, legal, tax or investment advice. Please seek professional advice based on your personal circumstances.