Kimi Lu
TLDR
The coming decade won’t be about making “coding” easier;
It’ll be about making software behave like bespoke media.
From shared products to personal programs
Era | How we “program” | Verifies against | User’s role |
|---|---|---|---|
1.0 | Hand-written code | Hard-coded rules & unit tests | Consumer of a rigid app |
2.0 | Train on data | Alignment with data-point distributions | Data point—usage refines a single global model |
3.0 | State goals; run agents | Goal achievement scored by automated + human verifiers | Director—each person coaches their own agent |
Software 1.0 and 2.0 scale by pooling everyone’s clicks or data; the product you and I use is largely the same.
Software 3.0 flips that logic: the model infrastructure improves in the background, but every run is tuned to one user’s goals, tools, and context. [Latent Space]
Why “vibe-coding” tools miss the bigger shift
New AI builders like Lovable let anyone prompt an LLM to spit out React components. It feels magical—finally a no-coder can assemble a Software 1.0 toy. But critics already warn that the code is brittle, verbose, and risky for core workflows; it’s a great demo, not a durable foundation. [Business Insider]
That nostalgia for GUI-draggers and auto-generated JSX is a diversion. The real leap is skipping deterministic code entirely and moving straight to goal-driven agents that verify themselves. In other words:
Let the old die. 3.0 doesn’t need prettier ways to author 1.0—it needs new ways to share, remix, and trust agents.
Software as content, not product
Think of a 3.0 agent as living knowledge: a bundle of skills, prompts, and reward logic that can be shared like a YouTube clip yet executed like software.
Dimension | 1.0 / 2.0 | 3.0 |
|---|---|---|
Form factor | Binary / SaaS endpoint | Goal + policy + verifier |
Customisation | Settings panel | Fork the agent, tweak two sentences, press play |
Distribution | App stores | Feed-style “remix & run” |
Creator archetype | Developer | Content-style builder (domain expert + narrative + prompt) |
Just as blogging became vlogging, authorship shifts once friction disappears. The next breakout “developers” will look more like TikTok creators—dropping automation templates, not tutorials.
What building & sharing will feel like
“Remix” replaces “install.” Clone an accounting agent, point it at your Stripe keys, change the tone to match your brand.
Reward scripts travel with the template. Safety rails and evaluation logic are part of the package—no more copy-pasting unit tests.
Analytics look like creator dashboards. You’ll track forks, successful goal completions, and audience growth, not monthly active users.
Long-tail uniqueness is an asset. Millions of niche runs teach the underlying verifier edge-cases without diluting personal customisation.
How to design for the content era
Optimise for forks, not features. Your best growth loop is users remixing your agent for their own context.
Bundle context lightly. Secrets, private docs, and corp data stay local; the shared template must remain stateless by default.
Treat rewards as creative voice. “Pass the integration tests” or “never send an off-brand email” is part of the promise, not an afterthought.
Build rollback into the workflow. Mistakes will happen; diffed policy logs and one-click reverts make experimentation safe.
Think distribution first. SEO, embeddings in social feeds, and share-to-run links will matter as much as code quality once did.
Closing thought
No-code movements have always tried to teach people to build traditional apps. Software 3.0 does something more radical: it turns software itself into a shareable medium, so the next generation of “developers” create by expressing goals, not by writing loops. When every ambitious professional can publish a living, self-verifying agent as easily as a TikTok, technology finally becomes personal at internet scale—and that’s the democratisation we’ve been promised for decades.
