Birgitta Böckeler's article on harness engineering at Thoughtworks distinguishes between 'guides' and 'sensors' for AI agents. Guides direct agent behavior proactively, while sensors catch and correct it after producing unsatisfactory outputs.
The article analyzes a discussion about GPT 5.5 in OpenClaw, concluding that complaints often target agent behavior, system prompts, or wrapper effects rather than the raw model quality itself. It emphasizes the importance of distinguishing between base model issues and prompt/runtime/product issues for effective optimization in AI workflows.
This content discusses implementing statistical guardrails for non-deterministic agents, which are systems where the same input can lead to distinct outputs across multiple runs. The purpose is to manage variability and ensure reliability in such agents.
