## Core Insight
When AI agents generate operational artifacts (interaction logs, code commits, review cycles, wage calculations), those artifacts become training data for fine-tuning the agents themselves. This creates a self-improving flywheel: better agents produce higher-quality operational data, which produces better fine-tuned models, which produce better agents. The company's daily operations are simultaneously its training pipeline.
Brian Roemmele's Zero-Human Company performs overnight LoRA fine-tuning on A100/H200 clusters using approximately 4 million high-quality examples derived entirely from internal operations. The training data includes agent interaction logs, code commits, review cycles, JouleWork thermodynamic wage calculations, and consensus outcomes.
## Key Principles
- Operational output = training input; every work cycle improves the next
- LoRA fine-tuning on domain-specific data creates models that outperform general-purpose models for that company's specific tasks
- The flywheel creates a proprietary data moat — competitors cannot replicate your model without your operational history
- Overnight fine-tuning cycles mean improvements compound daily
- Quality of training data matters: "high-quality examples" from structured processes (not raw dumps)
## Cross-Domain Applications
- **Entrepreneurship**: Operational fine-tuning creates a data moat that grows with usage. Each customer interaction makes the system harder to replicate. Connects to [[Data Lock-In as SaaS Survival Moat]] at the model level.
- **Knowledge Management**: Agent logs as organizational memory — the fine-tuned model captures institutional knowledge implicitly. Parallels [[Knowledge Infrastructure Compounding]] but at the model layer.
- **AI-Assisted Development**: Self-improving codebases where code review cycles train better code generators. The review process improves the thing being reviewed.
- **Career Strategy**: Operating fine-tuned models on proprietary data is a high-leverage skill that general "prompt engineering" doesn't capture.
## Connections
- [[Zero-Human Companies as Autonomous Agent Economy Pattern]] — The organizational context where this flywheel operates
- [[Multi-Agent Setup Iteration as Self-Improving Loop Requirement]] — Iteration at the configuration level; this note describes improvement at the model weight level
- [[LLM Model Selection for Specialized Tasks]] — Fine-tuning is the extreme version of model specialization: building task-specific models rather than selecting general ones
- [[Knowledge Infrastructure Compounding]] — Same compounding logic applied to model weights rather than documents
## Source
- [[You don't need Claude Code for OpenClaw]] (Brian Roemmele, February 2026)