Feedback on “12-Factor Agent”

[flingjie]

Great work summarizing the core principles — it’s a solid bridge between software deployment discipline and the emerging agent paradigm.

One suggestion: we could explore evolving this into a “12-Factor Agent 2.0”, shifting focus from deployment consistency to learning and data iteration.

Key ideas to consider:

• Treat data pipelines and experience management as first-class citizens — ensuring clean, versioned, and traceable datasets.

• Add a principle on feedback loops and reinforcement, connecting runtime metrics to model or policy updates.

• Introduce a notion of “token economy”, where interaction data itself becomes the foundation for continuous improvement.

These ideas are partly inspired by my LLM thinking — detail in From Tokens to Intelligence: How LLMs Learn, Think, and Act
.

Would love to discuss how we can evolve this framework toward a learning-centric agent architecture.