The End of Coding? Andrej Karpathy and the Rise of the “Loopy” AI Era
What happens when writing code is no longer the core activity in building software?
In a recent episode of the No Priors podcast, Andrej Karpathy describes a shift that is already underway: software development is moving from manual execution to AI-led orchestration.
This is not a distant trend. It is a change in how work gets done.
From Coding to Agentic Engineering
Traditional software development has been constrained by human effort—writing, reviewing, and maintaining code.
That constraint is shifting.
Karpathy describes a workflow where engineers no longer write most of the code themselves. Instead, they:
- Define tasks at a higher level
- Delegate execution to multiple AI agents
- Review and refine outputs
The limiting factor is no longer technical skill in a language or framework. It is the ability to:
Frame problems clearly and manage AI-driven execution.
This introduces a new capability layer inside organisations:
- From developers → to orchestrators
- From implementation → to direction and control
The Rise of Agents and the Decline of Traditional Applications
A second-order effect is emerging in how software is experienced.
Instead of interacting with multiple applications:
- Systems expose APIs
- Agents coordinate across them
- Users interact through natural language
Karpathy describes autonomous agents that can:
- Discover systems
- Interact with APIs
- Build working solutions without predefined interfaces
This suggests a shift away from:
- Fixed software products
toward - Dynamic, agent-assembled workflows
For businesses, this has implications for:
- Product design
- Integration strategy
- Customer experience
Moving from Interest to Capability
Many organisations are already experimenting with AI. Fewer have translated that into repeatable capability.
The gap is not access to tools. It is operational fluency.
For a structured approach to building this capability:
- AI Fluency fundamentals: https://claude.imbila.ai/ai-fluency.html
- The 4D Framework for applying AI in practice: https://claude.imbila.ai/4d-framework.html
These resources are designed to help teams move from isolated use cases to consistent, scalable application.
AutoResearch and the Automation of Discovery
One of the more significant developments is what Karpathy refers to as AutoResearch.
This involves systems that can:
- Generate hypotheses
- Write and test code
- Evaluate outcomes
- Implement improvements
In one example, an AI system identified optimisations that exceeded expert human tuning.
This points to a broader shift:
AI is not only automating execution. It is beginning to automate parts of the discovery process.
For knowledge-driven industries, this changes how advantage is created and sustained.
The Reality of “Jagged” Capability
Despite rapid progress, AI capability remains uneven.
- Highly effective in structured, measurable domains
- Less reliable in subjective or context-heavy scenarios
This creates what Karpathy describes as “jagged intelligence.”
For organisations, this reinforces two requirements:
- Strong human oversight
- Clear boundaries for where AI is applied
AI should be deployed where outcomes can be tested, measured, and iterated.
The Jevons Paradox and Software Demand
As the cost of producing software decreases, demand increases.
This aligns with the Jevons Paradox:
- Lower cost → more usage
- Easier creation → broader adoption
Rather than reducing the need for software, AI is likely to expand it significantly:
- More internal tools
- More automation layers
- More experimentation across functions
Digital Acceleration vs Physical Constraints
Karpathy makes a clear distinction:
- Digital systems will evolve rapidly
- Physical systems (robotics, infrastructure) will lag
The reason is straightforward:
- Software scales at near-zero marginal cost
- Physical systems do not
For business leaders, this suggests prioritisation:
- Focus AI efforts on digital workflows first
- Expect slower transformation in physical operations
A Shift in How Knowledge is Structured
One of the more subtle changes is how knowledge is created and shared.
Karpathy suggests that documentation will increasingly be written for machines first:
- Structured, machine-readable formats
- Designed for agents to interpret and act on
From there, AI systems translate that knowledge for humans.
This has implications for:
- Internal documentation
- Training and onboarding
- Knowledge management systems
What This Means for Business
Several practical implications emerge:
- Capability is shifting from execution to orchestration
The ability to direct AI systems will differentiate teams. - Agents will become a primary interface layer
Not replacing systems, but sitting across them. - Speed will no longer be a differentiator on its own
Clarity of intent and structure will matter more. - AI will amplify existing organisational strengths and weaknesses
Well-structured environments will scale. Poorly structured ones will struggle faster.
Closing Perspective
This is not the end of software development.
It is the end of manual coding as the central constraint.
In its place:
- Problems become the primary input
- Agents become the execution layer
- And orchestration becomes the core capability
Source: No Priors podcast episode with Andrej Karpathy