Harness engineering in practice: turning AI agent mistakes into rules of its environment (HDD)
The mistake-to-rule loop, guides and sensors, three layers of checks, how it fits with specs, and five first steps in your own project.
Blog: the author’s views and experience
Personal notes by Egor Urvanov on AI development, engineering management and communities. They reflect one author’s views and practice.
Approaches to working with a harness: the answer loop and review that fixes the original request
What to do when an answer misses: three paths and a repeated loop. How review finds gaps, why they go into the original request and the pipeline runs again. Based on Anthropic, OpenAI and Google guides.
A reconcile loop for specs: like Kubernetes, with the spec as desired state and code as actual
Spec-driven development with a reconcile loop: who changes the desired state and who brings the code to it, why green tests do not yet mean “correct”, and which six environment capabilities the agent needs to close the loop.
Specs by aggregate: RFC requirements and bounded contexts instead of a spec per feature
Spec-driven development without spec sprawl: how to keep specs in one aggregate dictionary so they do not multiply with every feature: requirement levels, the aggregate card and context boundaries.
AGENTS.md and CLAUDE.md, specs, rules and skills: what an agent should know about a project and where to keep it
Four kinds of persistent context for AI (AGENTS.md, CLAUDE.md, skills): the task spec, the project bible, rules and skills. What goes into each, when it enters the context, and what ChatGPT and Claude offer. Based on Anthropic, OpenAI, GitHub, Cursor and Cline docs.
Who to interview before buying AI tools: eight polar groups
How to pick 16 people from the report, a questionnaire and license lists, check team coverage and replacements, and decide what to ask them.