AI Agent Systems

This is a living note about my current understanding of AI agent systems. It is neither a finished theory nor a claim that the related projects are mature.

Starting point

I started using ChatGPT in 2022. Initially, it felt like a tutor and a smarter search engine. By 2024, I was bringing agents into projects, exploring MCP, LangChain, RAG and the ways AI could directly modify and deliver code.

My underlying view has not changed:

AI does not simply replace a profession. It redefines the people within it.

AI is a multiplier, but a multiplier cannot replace engineering ability. As agents take more actions, permissions, security boundaries, observability and recovery need careful design.

Current direction

Loong / loongclaw is one of my public agent projects. It explores a Rust foundation for domain-specific agents, focusing on:

• a secure, controlled execution foundation;
• composable long-running workflows;
• execution of compound tasks;
• collaboration between people and AI in real situations;
• a feedback loop from execution results to improvements in the system.

The repository is still evolving. It is better understood as an ongoing research and development direction than as a stable general-purpose framework.

Open questions

• How can tools be powerful while their permission boundaries remain understandable?
• What should memory retain, and what should it deliberately forget?
• How can state, retries and human confirmation become first-class features of long-running tasks?
• How much domain logic belongs in a specialized agent, and how much should remain in a shared runtime?
• How do we evaluate whether an agent has completed a task, rather than merely produced plausible output?

I will update this note as Loong evolves, as I work with MCP and as real products provide feedback.