I built ARGO, an open-source Agent Loop for traceable AI coding delivery

I kept running into the same failure mode in larger AI coding tasks: the agent could generate code quickly, but the decisions behind the requirement, architecture boundary, and acceptance criteria disappeared across sessions. When a test failed, the easy response was to patch the code. But sometimes the actual problem was an unexamined implementation assumption — or even an unclear business decision upstream. I built ARGO to make that delivery chain explicit. It is an open-source, workspace-installable Agent Loop framework for Cursor, GitHub Copilot, and OpenCode. ARGO combines two ideas: An ArchiMate-structured intent graph. Goals, capabilities, constraints, dependencies, work packages, and explicit acceptance test cases live in SystemArchitecture.json, which remains the canonical fact source. GraphRAG semantic retrieval plus a staged Agent Loop. It helps an agent navigate the intent graph, then drives work through business clarification → intent design → implementation design → coding/repair → two-level acceptance → archive. Each stage has a constrained responsibility. Intent design owns the acceptance meaning; implementation design creates stable boundaries and test entrypoints; coding changes production behavior but cannot rewrite frozen acceptance tests to manufacture a pass. The two acceptance levels are intentional: Does the code meet the implementation contract? Does the implementation meet the original business intent? If either answer is no, the GAP goes back to the stage that owns it instead of treating every failure as a code bug. There are two operating modes: human-in-the-loop for continuous approvals, and supervised autonomous mode for tasks with clear boundaries and executable verification. This is deliberately heavier than a prompt chain or lightweight spec workflow. I think it fits high-risk, multi-team, or auditable delivery; small, low-risk tasks may be faster without the full harness. GraphRAG semantic features also require environment setup (including Neo4j connectivity and a configured vector lifecycle). I would value critical feedback on: Which decisions do your agents lose between sessions? How do you determine whether a failure is intent, architecture, or implementation? Are the stage boundaries useful, or too rigid for real work?