Coding agents waste 80% of context on blind file dumps, audit reveals

Coding agents like Cursor and Claude aren't degrading—they're suffocating on bloated context windows before generating a single line of code. A developer who audited API logs and prompt payloads this week found that on repositories above 10,000 lines, agents recursively grep and dump 40+ files into context just to locate a single utility function, often hallucinating duplicate components when they can't find existing ones. Raw ingestion compounds the waste: agents load entire 2,000-line files to update five-line interfaces, while verbose test logs and MCP tool definitions consume 30,000 tokens before code generation even starts. Every session resets to zero project awareness, forcing agents to re-read the same files in a loop.

Once context hits 80 percent capacity with this noise, attention mechanisms collapse and code quality tanks. The productivity math inverts: save an hour typing, spend five hours untangling architectural spaghetti. Standard chunking RAG doesn't solve the problem because it can't capture logical structure—agents remain blind to how a codebase is organized until they've burned tokens on raw text. The developer calls for open-source agents that parse code into ASTs or graph databases to understand structure before wasting context, a capability no mainstream tool currently ships.