I built an autonomous coding agent that argues with itself before touching your code

Most AI coding tools have the same fundamental problem: they start writing code immediately. No verification, no pushback, no evidence. Just vibes. I spent 6 months building something different. MAESTRO runs an adversarial debate between two frontier AI models before executing anything. How it works You describe a task Two AI models — GPT-5.4 as proposer, Claude Sonnet as challenger — debate the approach. Each reads your actual files using a structural knowledge graph (Memgraph + Tree-sitter AST parsing). Each cites evidence. Each challenges the other's assumptions. The debate produces a specification: exact files to modify, constraints, test expectations, risk level You read it and approve it Only then does execution begin — inside an isolated Firecracker microVM Nothing commits until it passes 16 deterministic safety checks and a structured review The stack FastAPI backend, Temporal for workflow orchestration Firecracker microVMs for isolated execution Memgraph + Tree-sitter for stru