Key points
- Project at a glance: cc-haha is a desktop AI coding workbench rebuilt from leaked Claude Code source. In 40 days it accumulated 10.4K stars, 7.8K forks, 600+ commits (355 on desktop), and a single-day peak of 44 commits.
- Tech stack: Bun runtime, TypeScript, Tauri 2 desktop shell, React + Vite, React + Ink terminal UI, plus MCP and LSP protocols.
- Core capabilities: multi-session workspace, branch/worktree launching, right-side diff panel for AI edits, unified permission flow, multi-provider model routing, Computer Use (screenshot/click/type), H5 remote access via one-time tokens, IM integration (Telegram, Feishu, WeChat, DingTalk), scheduled tasks, and token usage analytics.
- Three paradigm shifts in AI coding: 1. *Copilot era (2021-2023)* — line-by-line autocomplete. 2. *Vibe Coding (2024-2025)* — conversational coding, coined by Andrej Karpathy. 3. *Agentic Engineering (2026+)* — humans orchestrate PEV (Plan-Execute-Verify) loops; AI agents perform multi-file edits, run tests, and self-correct. Karpathy announced vibe coding outdated in February 2026.
- Why "ordinary" builders can now ship: the bottleneck shifted from typing speed and syntax recall to problem definition, context organization, judgment, and validation loops. Andrew Ng predicts PM:Engineer ratios may invert from 1:4-6 to 2:1. Garry Tan noted 25% of YC W2025 startups generated 95% of code with AI.
- Why single models are not enough: context windows, knowledge cutoffs, hallucination, no real-world action, single-thread execution. cc-haha's remedy is an external capability matrix — browser/WebSearch fallback, Computer Use, MCP Skills, IM bridges, and multi-provider routing.
- Four-stage methodology observed in cc-haha: (1) minimum viable loop — get the leaked source running; (2) core experience loop — Tauri shell, sessions, permissions; (3) extended capability loop — Computer Use, IM, WebSearch; (4) quality and scale — CI/CD, tests, docs, governance.
- Quality gates accelerate iteration: PRs validated against real-model baselines, automated releases, and clean structure reduce bug-fix cost and unlock refactoring. The article proposes an Agentic Engineering quality pyramid: AI self-verification → static analysis → automated tests → human review.
- Responsibility shifts forward: humans now own goal definition, context curation, result judgment, and validation design, while AI handles execution and self-repair. Building becomes cheap; deciding what to build becomes scarce.
- Industry signals: Claw-code reached 100K stars in 24 hours; Lovable scaled from $1M to $100M ARR in 8 months; LinkedIn launched the Associate Product Builder role merging PM, design, and engineering; Naval Ravikant called vibe coding the new product management.
- cc-haha GitHub: https://github.com/NanmiCoder/cc-haha
- Andrej Karpathy on Agentic Engineering: https://twitter.com/karpathy/status/1886877799056142556
- Naval Ravikant on Vibe Coding as PM: https://twitter.com/naval/status/1886877799056142556
- LinkedIn APB Program: https://linkedin.com/apb
- Claw-code 100K Stars: https://github.com/ultraworkers/claw-code
- Agentic Engineering Guide: https://buttondown.com/verified/archive/the-2026-agentic-engineering-shift
Action checklist for solo builders
1. Start with the smallest closed loop — one idea → implementation → verification cycle. 2. Establish quality gates early, even if it is a single test. 3. Pair the model with browsers, MCP tools, and Computer Use so agents can act, not only chat. 4. Spend roughly 80% of effort on problem framing and result validation, 20% on execution delegated to AI. 5. Position yourself as an orchestrator: design constraints, specify intent, build PEV feedback cycles.
One-line takeaway
> The future barrier is not whether you can write code, but whether you can define problems, organize context, judge results, and build validation loops. cc-haha's 600+ commits in 40 days were not 600 typing sessions — they were 600 orchestrated PEV cycles.