Look at what your coding agents actually pull into context. For a single task, the agent greps the repo, runs a broad vector search, and loads dozens of files and chunks that merely resemble the request into the window before it writes a line. You pay for that twice. Once in tokens, for context the model never uses and once in precision, because the few facts that matter are now buried in lookalike text, and a window crowded with noise makes the model reason worse. The reflex may be to buy a longer context window, but a bigger window only buys room for more noise.
The fix is not more retrieval or a bigger window, it is structuring what you retrieve so the agent gets exactly what the task needs, which is cheaper and sharper at the same time.
Building the AI-DLC harness at Betsson we found that replacing "search the codebase" with a knowledge layer the agents query is the best way to get that right. It can be by reverse engineer the codebase, or to build a full knowledge base fit for Agent utilization. An agent picking up a task now pulls the exact typed slice it touches, the contract it must satisfy and the decision that governs it, instead of paying to read whatever text sits near the task in embedding space.







