Fixing Over-Engineering in LLM Coding Agents

Making a coding agent stop at the first solution that works, without cutting a single guard

Read this story for free: link

LLMs are now widely used for software engineering, and reasoning models added a thinking budget on top: the model reasons at length before it writes any code. The paper The Danger of Overthinking found that reasoning models overthink more than non-reasoning models on agentic tasks, and in my coding runs a reasoning model asked for a small change kept adding to it. People call the human version gold plating: looking at a task for so long that you deliver more than anyone asked for. Here is one example, a plain request I gave to an agent running Qwen/Qwen3.8-27B.

The same request answered two ways (Created by Fareed Khan)

As you can see, my request was small, yet the agent wrote 56 new lines, 30 of them code, including a DELETE that silently removes members. A developer needs one database constraint, and with our system the agent changed two lines of code in its first run.

Each extra line costs tokens and review time, and a DELETE like that can destroy data. Across my 34 tickets, a typical passing answer added 1.5 times the code the ticket needed. The software engineering approach I call proportional code targets that waste: before writing anything, the agent tries cheaper…