When Agents Learn From Work, the Data Wall Changes
Exhausting high-quality human text would not exhaust data for improving LLMs: agent work can produce new grounded experience, and selected experience can already improve later systems.
AI agent research notes
AgentCrucible Notes
Research notes on AI agent organization, agent self-evolution, evals, and project-grounded AI engineering.
Author
Boting Wang
Founder & CEO of AgentCrucible
Researcher of AI agent organization. Former software engineer at Microsoft and physics PhD at Southern Methodist University, writing about how useful agent systems can be organized, evaluated, and improved.
Focus
- AI agent organization
- agent self-evolution
- evals for agent systems
- AI engineering from real projects
Why Evals Are the Bottleneck for Useful Agent Systems
A project-grounded argument for why useful agent systems need eval infrastructure before they need more orchestration complexity.