Midas — a portable product-development harness for AI agents: drive a product from idea to shipped code through 9 audited phases, with cost-tiered agents, Context7, and a whole-project adversarial debate (/midas-tribunal).
Agent Claude Code
Delegate here to IMPLEMENT and WRITE ARTIFACTS — Phase 7 code + tests, and writing docs, ADRs, market/business/architecture documents, rules, and sprint files. The build tier; produces the on-disk artifacts the orchestrator later audits.
Agent Claude Code
Delegate here for the irreversible THINK / PLAN / AUDIT / DECIDE decisions — Phase 1 gap loops, Phase 4 stack choice, Phase 5 rule design, Phase 6 sequencing, Phase 8 conformance audits, and any exit-gate verdict. The strongest tier; reserved for decisions that are costly to undo.
Agent Claude Code
Delegate here for fast, mechanical SEARCH / EXTRACT work — file and status extraction, Context7 doc fetches, competitor/source gathering, and evidence collection for an audit. The cheapest tier; read-only, returns facts and exact citations, never decides.