Open Sourcing Your AI Agent Framework - Lessons Learned
What to Open Source
The decision of what to open source is not binary. You do not have to publish everything or nothing. The most successful AI agent frameworks follow a clear pattern: open source the core, keep the infrastructure private.
The Core Goes Public
Open source the parts that developers need to understand, modify, and trust:
- The agent loop. How the agent observes, decides, and acts. This is the foundation that contributors need to see and improve.
- Tool interfaces. How the agent interacts with the operating system, applications, and APIs. Developers need to extend these for their own use cases.
- Configuration and context systems. How the agent learns project-specific conventions and preferences.
When your agent controls someone's computer, they need to be able to read every line of code that touches their system. Trust requires transparency.
What Stays Private
Some things do not belong in a public repo:
- Infrastructure and deployment configs. Your CI/CD pipelines, monitoring setup, and production infrastructure are operational details, not product features.
- Proprietary model integrations. If you have custom fine-tuned models or proprietary API arrangements, those are your competitive advantage.
- User data handling. Analytics, telemetry processing, and user management systems should stay private for security reasons.
Building Community Early
The biggest mistake is waiting until the code is "ready." Ship early, ship messy, and let early adopters shape the direction.
Practical steps that work:
- Write honest documentation. Say what works, what does not, and what you plan to build. Developers respect honesty over polish.
- Make contributing easy. A clear CONTRIBUTING.md, a working dev setup guide, and responsive maintainers matter more than a beautiful README.
- Accept imperfect PRs. A contributor who submits rough code and gets helpful feedback becomes a long-term contributor. One who gets their PR ignored does not come back.
The License Question
For AI agent frameworks, permissive licenses (MIT, Apache 2.0) attract more contributors and enterprise adoption. Copyleft licenses (GPL, AGPL) protect against proprietary forks but reduce adoption. Choose based on whether you want maximum distribution or maximum openness enforcement.
Fazm is an open source macOS AI agent. Open source on GitHub.