Ai

Personal notes from Luo Fuli’s discussion of OpenClaw, and what it made me reconsider about agent orchestration, memory, cost, and where I should spend more attention as AI systems keep improving.

The ticket agent shipped and resolution speed went up by roughly an order of magnitude. That turned QA into the new bottleneck. This is what I learned about the other side of autonomous coding: Red/Green TDD for agents, rebuilding unit tests to be agent-writable, stopping the agent from cheating its own tests, and where hooks fit as the runtime enforcement layer.

LLMs got good enough that the hard problems shifted. I stopped writing about prompt engineering and started building autonomous coding agents. Here's what I learned moving from Claude Code to Pi, and why controlling an agent is harder than using one.

How adding observation masking to our existing LLM summarization reduced token costs, improved response latency, and eased rate limit pressure. Adapted from JetBrains research.

I wanted to understand how multi-hop RAG works and see if DSPy lives up to the hype. Built a demo with the MuSiQue dataset to explore iterative retrieval with synthesis-in-the-loop. Here's what I learned about adaptive reasoning and why it makes a real difference.