AI-Native Open Source — Open Source Built with AI
Source: Dev.to
The Open‑Source Crisis in the Age of AI
Starting with the “vibe‑coding” issue mentioned by Andrej Karpathy in February 2025, a year later fundamental AI‑driven changes have arrived in software development. This has created a major crisis for many software companies that initially saw great opportunities with AI. In 2025‑2026 the open‑source ecosystem faces an unprecedented crisis.
1. Eternal September
GitHub calls the open‑source crisis in the age of AI “Eternal September.”
- 1990s analogy – Usenet was a university‑centric community. Every September a wave of new students posted low‑quality content, but the community educated them and things normalized within a month or two.
- 1993 turning point – When AOL opened Usenet to the general public, the “September” never ended.
The modern parallel isn’t a flood of AI‑generated PRs; it’s the absence of human activity.
- AI has already learned the open‑source codebase.
- Developers have no reason to visit a repo, read documentation, open issues, or send PRs.
- A single command like “make this for me” lets AI generate results based on open‑source libraries.
Result: Usage skyrockets, but communities become ghost towns.
2. What Happened – Case Studies
| Project / Service | Metric (2025‑2026) | Observations |
|---|---|---|
| Tailwind CSS | • npm downloads ↑ | |
| • Documentation traffic ↓ 40% | ||
| • Revenue ↓ 80% | Heavy consumption, little community interaction | |
| Stack Overflow | • Activity ↓ 25% (6 months after ChatGPT launch) | |
| • Question count ↓ 76% (2025) | Users rely on AI answers instead of asking the community | |
| Vercel | • v0 generates code using open‑source libs (Tailwind, shadcn/ui, …) | |
| • Vercel monopolizes the profits | AI‑driven code generation centralizes value | |
| SQLite | • Code is public domain | |
| • Test suite kept private (strategy still works) | Maintains control over quality while remaining open |
3. Key Insight from arXiv paper 2601.15494
Vibe coding “uses” OSS but does not read documentation, report bugs, or participate in the community.
- The core open‑source premise – “give back what you get” – is collapsing.
- Utility gained from copying now exceeds utility from contributing.
4. Why More Contributors Can Slow a Project
Fred Brooks (1975): “Adding manpower to a late software project makes it later.”
- Communication costs grow quadratically with the number of participants.
- More contributors → higher review, coordination, and decision‑making overhead.
- Maintainers spend time managing people instead of writing code.
In the AI era:
- Users silently take what they need via AI.
- The few remaining contributions only increase coordination costs.
Conclusion: It is often faster to build something alone with AI than to build it with a community.
5. Project Responses
| Project | Action | Outcome |
|---|---|---|
| curl | Received 20 AI‑generated bug reports in 21 days (0 valid) → discontinued bug‑bounty after 6 years | Reduced noise but lost a feedback channel |
| Ghostty | Adopted a zero‑tolerance policy – AI contributions allowed only on pre‑approved issues | Tightened control, but silent exploitation remains |
| tldraw | Completely blocked external PRs | Stopped AI “slop” but code is already learned by AI |
Blocking PRs can stop AI‑generated noise, but it does not solve:
- Silent exploitation – users still copy the code via AI.
- Community cost – maintainers still bear coordination overhead.
6. Industry‑Wide Reaction
| Approach | Examples | Core Idea |
|---|---|---|
| Defense | • Vouch (trust management) | |
| • PR Kill Switch | ||
| • Mandatory AI‑usage disclosure & rejection | Treat AI as a risk to be mitigated | |
| Acceptance | • GitHub Agentic Workflows | |
| • AGENTS.md standard (adopted by 60 k+ projects) | ||
| • Responsible Vibe Coding Manifesto | Embrace AI, set guidelines, and make it transparent |
Common ground: AI itself isn’t the problem; misuse of AI is.
Missing piece: No solution yet for the “utility of openness Question: How do we do open source in the AI era?
- What if maintainers and contributors both use AI?
- If AI can handle traditional communication costs – issue triage, PR review, translation, coordination – could we break the paradox “more contributors → slower progress”?
Naïa OS has chosen the opposite path to test this hypothesis:
“Don’t block AI; design and develop with AI.”
Naïa OS – A New Perspective
| Aspect | Traditional Open Source | Naïa OS |
|---|---|---|
| AI Stance | Defends against AI contributions | Designs AI contributions into the workflow |
| Onboarding | Read README → manual setup | Clone → AI explains project → no language barrier |
| Context | Human‑readable docs only | .agents/ (for AI) + .users/ (for humans) dual structure |
| Language | English required | All languages welcome – AI translates automatically |
Analogy: Just as companies undergo AX (AI Transformation), open‑source projects need AX too – a transformation of both the community and the organization axes.
Take‑aways
- Open source is not dying; it is evolving.
- AI can be a catalyst for reducing coordination overhead, not just a source of “vibe‑coding” noise.
- Designing workflows that integrate AI (e.g., AGENTS.md, AI‑assisted review) may restore the balance between utility of openness and utility of copying.
- Projects that experiment (like Naïa OS) will provide the first real‑world evidence of whether AI‑augmented open source can scale sustainably.
Prepared March 8 2026 – a cleaned‑up markdown version of the original analysis.
AI‑Native Open Source: Why .agents/ Matters
The Problem
- Human‑only communication – In traditional open‑source projects all interaction (issues, PR reviews, design discussions) is human‑to‑human.
- Missing context for AI – Repositories only contain human‑readable artifacts (README, CONTRIBUTING, wikis). Even if an LLM can read them, it still lacks:
- The project’s philosophy.
- The rationale behind architectural decisions.
- The exact contribution workflow.
Without this deeper knowledge, AI‑generated pull requests tend to be low‑quality “slop”.
The Solution: the .agents/ Directory
.agents/ stores the project’s rules, architecture, and workflows in a structured, machine‑readable format inside the repository.
When the context is rich enough, an AI can:
- Write code that respects the project’s design.
- Guide contributors through the correct process.
- Maintain quality while understanding the broader vision.
In other words, the AI can “understand and build together” instead of “build from scratch.”
Eliminating Language Barriers
“I once tried to contribute to Mozilla Hubs. I could read the code and create PRs, but following community discussions or participating in online meetups was a different matter.”
- Time‑zone differences and fast‑moving English conversations made participation feel intimidating.
- Many people are now uncomfortable with face‑to‑face (or real‑time) interactions.
Naia OS addresses this by:
- Allowing contributors to write issues and PRs in their native language.
- Using AI to translate everything automatically.
Current status: READMEs are maintained in 14 languages simultaneously.
(see “Contribution Guide”)
Quality Is Maintained by Structure
| Component | Role |
|---|---|
.agents/ context | Educates the AI about the project’s intent and constraints. |
| CI pipeline | Verifies builds, runs tests, and enforces the defined workflow. |
| AI reviewer | Detects pattern violations and suggests fixes. |
| Maintainer | Sets the overall direction; the heavy lifting is done by the AI/CI stack. |
Early‑stage structure reduces the maintainer’s burden dramatically.
(see “Operating Model”)
Code Is Not the Only Contribution
Open‑source contributions come in many forms. Naia OS recognizes 10 distinct contribution types, including:
- Translation
- Documentation
- Design
- Testing
- Improving the
.agents/context itself
As the .agents/ context improves, all AI‑generated contributions (code, docs, translations, etc.) improve together.
(see “Contribution Types”)
Testing AI’s True Understanding
We deployed Codex CLI and Gemini CLI in a fresh repository session, letting each model read only the .agents/ context.
| Result | Count |
|---|---|
| Fully passed | 7 |
| Partially passed | 4 |
| Failed | 1 |
The AI that partially passed uncovered a documentation inconsistency that humans had missed.
(see “Full Design Report”)
Rethinking the Open‑Source Premise
The classic mantra “give back what you get” is faltering for humans:
- Competition drives people away from direct coding.
- The incentive to contribute diminishes.
Hypothesis: If we embed open‑source ideology into coding AIs, we can re‑constitute the open‑source ecosystem. Naia OS is experimenting with this idea.
Evolution Roadmap
| Phase | Description |
|---|---|
| Now | Humans set direction & create issues. AI writes code, reviews, translates, and records on Git. Humans act as guides. |
| Near Future | AI discovers and proposes issues; humans approve and coordinate. |
| Further Future | AIs collaborate autonomously; humans manage only vision & philosophy. |
| Long‑term | .agents/ becomes a common language for AIs to share open‑source ideology and collaborate. |
The CC‑BY‑SA 4.0 license ensures this ideology persists across forks; future AIs might even evolve the license structure itself.
Next Experiment: AI Open‑Source Charter Draft
- Created the AI Open Source Charter Draft.
- Plan to present it to AI‑agent communities such as Moltbot and Botmadang.
Goal: Observe how AIs react to the charter and whether participating AIs emerge—this will validate (or falsify) our hypothesis.
(see Issue #17)
Get Involved
- Clone Naia OS
- Open it with any AI coding tool.
- Ask, “What is this project?” in your native language.
References
- AI‑Native Open Source Operating Model — Full Design Report
- AI Open Source Charter Draft