OpenClaw Practical Evolution Theory Master Class

About This Activity

A master class on the practical evolution theory behind OpenClaw, held in Shenzhen the day before the Tencent OpenClaw hackathon. The session worked through how an agent framework evolves a solution rather than executing a fixed plan: how candidate actions are generated, how they are scored against the environment, and where the loop stops being search and starts being engineering. The material sat closer to systems design than to prompt writing. Most of the time went on the unglamorous parts of an agent harness, which is where the failure modes live: tool definitions precise enough that the model cannot misread them, state carried between steps without silently growing, checkpoints where a human decides whether the run continues, and what a sensible budget looks like before a loop is allowed to run unattended. Useful preparation for the competition the following day, and a clear picture of how the Chinese agent tooling community frames these problems in Mandarin, which is not always how the same problems are framed in English-language writing on agents.