I noticed my AI agent getting dumber the bigger my project got.
I started with Claude Code. First few days were magic. Then around week two — the agent started hallucinating functions that didn't exist. It got confused about what I was asking. More and more bugs. Every new feature harder than the last. I was spending more time fixing the agent's output than writing code myself.
I kept blaming the AI. But it wasn't the AI losing capability. It was my codebase losing structure.
Here's what was actually happening: same function names with different purposes scattered across files. Unrelated code dumped in the same folder. Dependencies tangling into spaghetti. When the agent searched my project, twenty conflicting results came back — and it picked the wrong one. Every session made the mess worse. Every mess made the next session harder. Eventually even the agent struggled to implement new features in its own codebase.
I looked at tools like Spec Kit — plan architecture first, then let AI implement. But that's not how I work. I prototype fast, chat with the agent, share ideas, follow inspiration. That creative flow is what makes AI agents powerful — and it's exactly what produces messy structure. AI agents can't focus on the big picture and small details at the same time.
What I built
A real-time feedback sensor. It watches the codebase — every file, every dependency — visualized as a live interactive treemap that updates as the agent writes code. 14 quality dimensions graded A through F.
For the demo I gave Claude Code 15 detailed step-by-step instructions to build a FastAPI API. Explicit module boundaries, explicit file separation. Five minutes later: Grade D. Cohesion F. 25% dead code. Even with careful instructions.
It also runs as an MCP server — the AI agent queries the grades mid-session, sees what degraded, and self-corrects. The feedback loop closes.
Pure Rust, single binary, 23 languages via tree-sitter, MIT licensed. Happy to answer any questions.
davej32•2h ago
I noticed my AI agent getting dumber the bigger my project got.
I started with Claude Code. First few days were magic. Then around week two — the agent started hallucinating functions that didn't exist. It got confused about what I was asking. More and more bugs. Every new feature harder than the last. I was spending more time fixing the agent's output than writing code myself.
I kept blaming the AI. But it wasn't the AI losing capability. It was my codebase losing structure.
Here's what was actually happening: same function names with different purposes scattered across files. Unrelated code dumped in the same folder. Dependencies tangling into spaghetti. When the agent searched my project, twenty conflicting results came back — and it picked the wrong one. Every session made the mess worse. Every mess made the next session harder. Eventually even the agent struggled to implement new features in its own codebase.
I looked at tools like Spec Kit — plan architecture first, then let AI implement. But that's not how I work. I prototype fast, chat with the agent, share ideas, follow inspiration. That creative flow is what makes AI agents powerful — and it's exactly what produces messy structure. AI agents can't focus on the big picture and small details at the same time.
What I built
A real-time feedback sensor. It watches the codebase — every file, every dependency — visualized as a live interactive treemap that updates as the agent writes code. 14 quality dimensions graded A through F.
For the demo I gave Claude Code 15 detailed step-by-step instructions to build a FastAPI API. Explicit module boundaries, explicit file separation. Five minutes later: Grade D. Cohesion F. 25% dead code. Even with careful instructions.
It also runs as an MCP server — the AI agent queries the grades mid-session, sees what degraded, and self-corrects. The feedback loop closes.
Pure Rust, single binary, 23 languages via tree-sitter, MIT licensed. Happy to answer any questions.