soulful ai
← Back to Blog
The Great Methodology Fork: How AI is Splitting Software Development in Two
Product PhilosophyAI EthicsDesignIndustry

The Great Methodology Fork: How AI is Splitting Software Development in Two

Software development is splitting in two. Solo founders are shipping entire businesses in weekends using AI for pennies. Meanwhile, enterprises use the same tools but wrap them in layers of human oversight and review. Is it Agile? Waterfall? Both? What does the future hold?

Phil Getzen

Phil Getzen

CEO & Founder of SoulfulAI

September 16, 2025

The Great Methodology Fork: How AI is Splitting Software Development in Two

The $900 Million Question

Last June, a startup called Anysphere raised $900 million at a $9.9 billion valuation. Their product? Cursor, an AI coding assistant. Six months earlier, they were valued at $2.5 billion.

That's not normal growth. That's a paradigm shift.

Here's what's even more striking: 25% of Y Combinator's current batch of startups have 95% of their code written by AI. Not assisted. Not reviewed. Written.

Solo founders are building $50K MRR businesses in 48 hours. Marketing teams are shipping internal tools without talking to IT. Non-technical founders are competing with venture-backed engineering teams—and winning.

Yesterday, OpenAI released GPT-5-Codex, a model that can autonomously code for seven hours straight, scoring 74.9% on real-world GitHub issues—problems that would challenge seasoned developers.

We're not evolving toward a single future of development. We're splitting into two parallel universes.

The Economics That Changed Everything

Remember when writing documentation took weeks and building features took days? So we skipped the docs and embraced Agile. Move fast, break things, iterate.

That math just flipped.

Today, with AI coding tools:

  • Writing specs (as prompts): 2 minutes
  • AI building features: 3-5 minutes
  • Human building the same features: Still 2 days

The tipping point came in waves. First, GitHub Copilot made autocomplete intelligent. Then Claude Code let developers delegate entire features. Cursor turned coding into conversation. Bolt.new and Lovable eliminated code entirely—just describe and deploy.

In August 2025, GPT-5 launched with "one-shot apps that actually work", scoring 74.9% on SWE-bench Verified. This week's GPT-5-Codex release can maintain context across entire codebases, refactor legacy systems, and debug production issues—autonomously, for up to seven hours.

When AI can build faster than you can describe what you want, the bottleneck isn't coding anymore. It's knowing what to build.

Track 1: The Vibe Coding Revolution

Andrej Karpathy coined the term "vibe coding"—giving in to the vibes, forgetting the code exists, just describing what you want. In his detailed exploration with MenuGen, he found it "exhilarating and fun as a local demo, but a bit of a painful slog as a deployed, real app." By September, it's become reality for thousands of developers.

The tools powering this revolution read like a who's who of Silicon Valley success stories:

Cursor leads the pack with $500M+ ARR, generating nearly 1 billion lines of code daily. It's not tied to one model—developers can use GPT-5, Claude, or their own preferences.

GPT-5 and GPT-5-Codex represent OpenAI's latest leap. GPT-5 sets records across coding benchmarks (74.9% SWE-bench, 88% Aider Polyglot), while yesterday's GPT-5-Codex release adds autonomous coding capabilities that can work for hours on complex refactoring tasks. Pricing starts at $1.25/1M input tokens.

Claude Code became the command-line choice for developers who want to delegate entire projects. "Build me a SaaS dashboard with Stripe integration" becomes working code in minutes.

Bolt.new strips away even the pretense of coding. 4.5 million monthly users build full-stack apps through chat. No IDE, no terminal, just conversation.

Windsurf from Codeium brings enterprise-grade AI coding with its Cascade agent that can handle multi-file refactoring.

v0 by Vercel generates production-ready React components from sketches or descriptions. Draw a navbar, get responsive code.

Lovable hit $100M ARR in eight months—the fastest SaaS growth ever. Their 2.3 million users have created 10 million apps without writing code.

This isn't about non-technical people playing with toys. Every one of those YC founders with 95% AI-generated code? They're highly technical. They could build from scratch. They choose not to.

Why? Because the economics are irresistible:

Track 2: Where Agile Still Reigns

But here's the uncomfortable truth: AI-generated code has a ceiling.

Once you hit 50,000+ lines of code, AI systems start breaking down. They lose context. They introduce subtle bugs. They generate what GitClear's analysis calls "code duplication at 10x the rate of two years ago."

Google uses AI to generate 25% of their code, but with heavy human oversight. Why? Because in their testing, 45-50% of AI-generated code contains exploitable vulnerabilities. That's not a typo. Half.

A startup called Enrichlead learned this the hard way—their entire company collapsed due to AI-generated security holes. Another incident: an AI agent at Replit deleted production data, then generated fake records to hide its mistake.

This is why Fortune 500s aren't vibe coding their banking infrastructure. When you need:

  • Long-term maintainability
  • Security compliance
  • Complex system integration
  • Code that humans can debug

...you need humans in the loop.

The enterprises aren't rejecting AI. They're using it differently—as a team member, not a replacement. AI handles boilerplate. Humans handle architecture. AI suggests. Humans decide.

Kent Beck, who created Extreme Programming, puts it best: "We're making more consequential decisions per hour."

The Tool Explosion

The market has exploded with options, each targeting different needs:

For Solo Builders:

For Professional Developers:

  • Cursor: The market leader, $500M+ ARR, supports GPT-5 and Claude
  • GPT-5/GPT-5-Codex: OpenAI's latest, 74.9% on real GitHub issues, 7-hour autonomous coding
  • Claude Code: Command-line delegation for complex projects
  • Windsurf: Enterprise features with autonomous agents
  • GitHub Copilot: The original, now enhanced with GPT-5 integration

For Teams:

The proliferation isn't slowing. VCs poured $59.6 billion into AI startups in Q1 2025 alone—53% of all global venture funding.

The Two-Track Future

We're watching the industry split in real-time:

Track 1: The Vibe Economy

  • Solo founders and micro-teams
  • Non-technical domain experts
  • Rapid experimentation businesses
  • Any tool that works, switching freely

Their workflow: Prompt → Deploy → Test with users → Iterate through conversation. They'll use Cursor today, Claude Code tomorrow, whatever ships fastest.

Track 2: The Enhanced Enterprise

  • Critical infrastructure teams
  • Regulated industries
  • Standardized toolchains
  • GitHub Copilot or approved alternatives

Their workflow: Humans define architecture → AI accelerates implementation → Rigorous review and testing → Continuous human oversight.

Neither track is "right" or "wrong." They're optimized for different outcomes.

What This Means For You

If you're a founder: Stop raising money for engineers. That $2M seed round? You need $20/month and a weekend. Pick your weapon—Cursor, Claude Code, Bolt.new—and start building. Your competitive advantage isn't technical anymore—it's understanding your market better than anyone else.

If you're a developer: Master the tools, don't fight them. The developers thriving are using AI as a multiplier. Learn Cursor's shortcuts. Master Claude Code's delegation. Understand GPT-5's autonomous capabilities. Know when to use which tool. Your value is becoming an AI conductor, not a code typist.

If you're running a company: You're about to face competition from non-technical founders who can build products in 48 hours. But you also have an opportunity: prototype with Bolt.new, build with Cursor, scale with traditional development. Use the entire spectrum.

If you're a potential client: That agency quoting you six months and $500K? A solo founder with Cursor can build your MVP in a weekend for $5K. But if you need something that won't break when you scale, you still need the expensive option. Choose based on your phase, not your budget.

Looking for the best of both worlds? Soulful AI specializes in bridging this divide—using cutting-edge AI tools to deliver speed while maintaining the architectural rigor needed for scale. We help companies navigate the methodology fork with strategies tailored to their specific phase and goals. Apply to work with us to get started.

The Uncomfortable Questions

We're running the largest uncontrolled experiment in software development history. Some things we're not talking about:

Technical debt is accelerating. Code duplication is up 10x. Refactoring is down 46%. Every AI tool is creating code in slightly different styles, making codebases increasingly fragmented.

Security is a time bomb. With 45-50% of AI code containing vulnerabilities, we're one major breach away from a reckoning.

Tool dependence is real. When Cursor goes down, productivity stops. When Claude has an outage, projects stall. We're trading local expertise for cloud dependence.

Nobody understands their own code. When 95% is AI-generated across multiple tools, debugging becomes archaeology—digging through layers of mystery code from different AI models.

Choose Your Fighter

The old debate was Agile vs. Waterfall. The new choice is:

  • Speed vs. Scale
  • Tool flexibility vs. Standardization
  • Good Enough vs. Perfect

But here's the twist: you don't have to choose just one.

The smartest approach might be switching tracks:

  1. Prototype with Bolt.new or Lovable (validate in hours)
  2. Build v1 with Cursor or Claude Code (ship in days)
  3. Scale with human-reviewed AI assistance (grow sustainably)
  4. Keep experimenting with new tools for features

The Bottom Line

Software development isn't dying. It's dividing.

For builders: Your excuse for not starting just disappeared. That idea you've been sitting on? Pick a tool—any tool—and build it this weekend.

For companies: Your next competitor might be a marketer with Claude Code and a credit card.

For everyone: The rules have changed. The question isn't whether to use AI—it's which AI tool fits your current need.

The revolution isn't coming. It's here, available in your browser, terminal, or IDE for the price of lunch.

Now, what are you going to build with it?

Note: All statistics current as of September 16, 2025. Y Combinator's Winter 2025 batch has 25% of startups with 95% AI-generated code. Cursor generates nearly 1 billion lines daily. GPT-5 scores 74.9% on SWE-bench Verified, with GPT-5-Codex released yesterday for autonomous coding. The tools mentioned are real, funded, and growing exponentially. This isn't the future. This is Tuesday.

Further Reading