A solo developer has built an open-source AI decision framework while documenting an instruction architecture for managing coding agents across complex monorepos.
Gábor Mészáros created SageCompass, combining LangGraph's stateful multi-agent workflows with Drupal for RAG content management and Gradio for UI. The project evaluates AI business ideas through structured stages: Problem Framing, Goals/KPIs, Feasibility Assessment, and Decision Synthesis.
The real contribution is the instruction maturity model. After two months of iteration without prior Python expertise, Mészáros found that different components needed different levels of guidance. LangGraph had tight contracts and clear boundaries. Drupal required codifying 17 years of domain knowledge. Gradio was still being learned.
The resulting framework spans six levels: L0 (no instruction file), L1 (basic tracked file), L2 (explicit constraints using RFC 2119 keywords), L3 (external references and modular docs), L4 (path-scoped rule loading), L5 (structural discipline with changelogs), and L6 (dynamic context with Model Context Protocol integration).
This matters because enterprise teams face the same problem at scale. A payment processing service needs different agent instructions than a prototype feature. Path-scoped loading (L4) prevents coding agents from applying API rules to test files. Structural discipline (L5) makes instruction changes auditable.
The approach emphasizes real projects over tutorials. Mészáros iterated from basic LangChain to agentic platform, refactoring for stability after early "hammer/nail" pitfalls. The monorepo enforces architecture via AGENTS.md, CLAUDE.md, and langgraph-python-architecture-principles.md.
LangGraph enables the stateful, graph-based workflows with cycles, memory, and human-in-the-loop capabilities. The latest changelog entry adds TraceEvent DTO for event deduplication, though timestamps weren't specified.
For CTOs evaluating agentic workflows: the instruction maturity model provides a framework for thinking about agent guardrails. The trade-off is upfront documentation cost versus reduced refactoring time. History suggests agentic stacks demand strong boundaries to avoid non-determinism.