Agent Harness — ECC

After working extensively with AI coding agents, I noticed that a large fraction of the errors and inefficiencies weren't capability failures — they were context failures. The agent didn't know what approach had already been tried, didn't remember that a particular pattern had caused problems before, or didn't have access to the non-obvious constraints that a human engineer carries implicitly. ECC is a harness for solving those problems systematically.

The skill system allows common agent behaviours to be codified and reused across sessions. Instead of re-specifying how to handle a database migration or how to structure a PR, you reference a skill by name and the agent has a complete, validated procedure to follow. This is analogous to runbooks in operations — the skill doesn't replace judgment, but it gives the agent a proven starting point.

The instinct layer is different from skills: instincts are patterns the agent should apply automatically without being instructed. Things like "always check for environment variable exposure before committing", "prefer editing existing files over creating new ones", or "read before writing". These are the accumulated lessons from many sessions distilled into standing instructions. The memory system persists findings and decisions across sessions so context doesn't reset to zero each time.