github.com• Feb 27, 2026• 1 min read
Show HN: Patterns for coordinating AI agents on real software projectsI'm a former chef with no CS degree, running 3 AI agents (PM, Builder, Auditor) building a D&D 3.5e combat engine. Every pattern in this repo was born from a specific failure.The failure catalog is the real value. Examples:- "Ghost Target" (pattern #3): Agent confidently "fixed" a bug that didn't exist. Now we verify gaps before writing code. - "Write-Only Field" (pattern #8): Agent wrote data at character creation that nothing ever read at runtime. 17 fields exposed. Now every data write requires a consume-site proof chain. - "Parallel Drift" (pattern #9): Two code paths computing the same attack logic diverged by 21 modifiers. Now every resolver change requires parity verification across all paths. - "Coverage Theater" (pattern #12): 30+ rows in the coverage map went stale. Tests passed but tracked nothing. Now coverage updates are mandatory in every debrief. - "Governance Consumption Failure" (pattern #15): PM wrote rules. Nobody read them. Process docs had no enforcement mechanism.