AI Coding: From Flow State to Review Mode
AI now generates 90% of code, killing hand-coding joy but demanding deeper code review skills as costs rise—stick to TypeScript/Python, embrace local models, build/review hybrids.
Agentic Coding Reshapes Developer Workflow
Maximilian Schwarzmüller shares that 90% of his code is now AI-generated via agentic workflows, where he defines base types/interfaces, prompts the AI, reviews output, and iterates fixes. This eliminates the 'flow state' joy of hand-typing: "the pure process of typing of getting into that flow state of writing code was a lot of fun... And I lost that." Instead, work shifts to spec-writing, code review, and babysitting AI output—especially painful for QA engineers facing tripled expectations and infinite AI slop to test.
He predicts hybrid approaches: usage-based pricing (e.g., GitHub Copilot's shift from subscriptions to per-token credits) will make hand-coding with basic autocompletion cheaper than full agents for simple tasks. Compute constraints from agentic token bloat ensure this persists short-term, reviving coding as a cost-saving skill: "writing something by hand maybe with AI powered autocompletion... may be more cost effective." Long-term, prices drop with supply, but reviewing AI code remains key to avoid degrading software quality.
Local models like Ollama and LM Studio shine for non-agentic tasks (text analysis, private data)—not yet viable on his 4-year-old M1 MacBook Pro for complex coding, but fine-tunes for niches could enable on-device task-specific agents, keeping sensitive data in-house.
Tool Shifts and Practical Alternatives
GitHub Copilot's usage-based pivot kills subscription value (prepay credits only), pushing to Cursor (VS Code fork, subscription-based), Codeium Cloud, or OpenCode Black. All likely follow suit eventually due to inference costs.
Backend stacks: Node.js (Bun runtime + Hono framework for speed) or Python/FastAPI—pick based on preference, as AI excels out-of-box in both. TypeScript edges out for types aiding AI accuracy.
Post-TypeScript/React: Build demo projects (AI-assisted but comprehend output), then Next.js or TanStack Start (meta-frameworks), or React Native for mobile. Avoid vibe-coding; understand languages deeply to steer/review: "in order to truly understand it, in order to be able to review code and instruct the AI properly, you nonetheless must understand the programming language."
Missing tools: Robust agent memory and universal CLIs/APIs for services—gaps ripe for disruption, as agent-friendly software wins.
Career Adaptation in AI Job Market
Coding fundamentals endure: AI broadens shallow knowledge via patient Q&A but forgets fast without deep dives (docs, hands-on builds). Juniors learning sans AI build vital bases.
Layoffs blend overhiring corrections with AI excuses; US Indeed data shows steady developer job rises post-pandemic lows, unlikely to peak at bubble highs but stabilizing. Companies need humans to leverage AI, preferring seniors—but juniors grow via review roles.
No pivot to ML engineering (low demand for trainers; focus on AI-users). QA/dev roles evolve to efficiency boosters via specs/reviews, not replacement. Overwhelm is normal—skip hype (e.g., MCP servers faded); monthly catch-ups suffice: ask GPT for updates.
">Quote: "I'm not switching profession. I'm not moving away. But... that flow state is gone."
Quote: "AI can generate an infinite amount of stuff... but you as a human you have a limited amount of time to review stuff."
Quote: "You're not left behind if you're not up to date all the time... it's all changing so quickly."
Quote: "The knowledge is getting way broader... but it's a very shallow knowledge and you forget it quickly."
Key Takeaways
- Define types/interfaces upfront, prompt AI agents, then rigorously review/fix output for 90% code generation.
- Switch from Copilot to Cursor or Codeium amid usage pricing; expect all tools to follow.
- Favor TypeScript/Python backends (Hono/Bun or FastAPI); types boost AI reliability.
- Use local Ollama/LM Studio for private/text tasks; await fine-tunes for agentic viability.
- Build/review hybrids beat vibe-coding; hand-code simple tasks to cut token costs.
- Learn deeply post-basics (Next.js after React/TS); understanding trumps shallow AI queries.
- Ignore job panic—review roles persist; catch up monthly, not daily.
- Target agent-tool gaps: memory, CLIs/APIs for services.