AI coding agents are great at generating code, but terrible at respecting architecture.
We kept getting code that looked correct but didn’t type-check, had hallucinated imports,
or ignored project boundaries.
AgentGuard is an open-source quality-assurance engine that sits between an AI coding agent and the LLM. It enforces a top-down generation pipeline:
Skeleton → Contracts → Wiring → Logic
Each stage constrains the next, and we validate syntax, imports, linting and types before any human sees the code. The LLM also self-reviews its output against explicit criteria.
AgentGuard supports MCP (Claude Desktop, Cursor, Windsurf, Cline) and tracks token cost per model and run.
Python 3.11+, MIT licensed.