From If/Else to Goal‑Oriented Agents: How LLMs Are Shaping Software 3.0

Software Evolution

Software 1.0 is "people write code, machines execute". Developers use explicit if/else statements to tell a computer each step, akin to teaching an obedient but non‑thinking assistant.

Software 2.0 shifts to "people provide samples, machines learn". By feeding data to train models, the system discovers how to act on its own. The process is somewhat a black box, yet it yields strong results, similar to an apprentice that mimics well.

Software 3.0 advances to "people state goals, machines understand and execute". Natural language prompts replace explicit code, allowing the AI to decide the implementation. This is like commanding a clever but occasionally errant assistant—no longer writing detailed code, just issuing high‑level directives.

LLM as Infrastructure

The author likens large language models (LLMs) to a new form of electricity, a basic public utility. Building an LLM requires massive capital and top‑tier talent, comparable to constructing a power plant or a semiconductor fab.

A more fitting analogy is that an LLM functions as an operating system and an expanding software ecosystem. Its capabilities depend not only on the model itself but also on the surrounding development and deployment tools.

Agent Engineering

The "Autonomy Slider" is presented as essential product engineering for maximizing an agent’s output. By employing multi‑round prompts, verification steps, and controllable autonomy—similar to autonomous‑driving levels (L1‑L5)—humans can intervene when needed.

Agents should be “tethered”: use concise, specific prompts and enforce acceptance logic on outputs. Broad prompts like “teach me physics” can cause the agent to wander aimlessly, delivering unsatisfactory answers.

Rather than debating when AGI will arrive, the focus should shift to improving product and technical engineering quality, making the autonomy slider easier for users.

Programming Paradigm Shift

The term "Vibe Coding" first appeared in a February tweet by Karpathy. He argues that AI engineers no longer need 3‑5 years of deep domain study; those skilled in prompt design, tool composition, agent coordination, and system verification will become the new key players, moving from traditional if/else implementations to natural‑language guidance.

Engineering Challenges

While programming becomes simpler, deployment remains complex: identity security, access authentication, payment verification, observability, and stability all demand robust infrastructure and architectural expertise.

Traditional infrastructure serves human‑centric applications; AI infrastructure should be purpose‑built for AI workloads. This creates opportunities for "AI Infra Builders" to shape the next generation of AI‑centric platforms.