10 things I learned from burning myself out with AI coding agents – Ars Technica

with LLMs, context is everything, and in language, context changes meaning. Take the word “bank” and add the words “river” or “central” in front of it, and see how the meaning changes. In a way, words act as addresses that unlock the semantic relationships encoded in a neural network. So if you put “checkerboard” and “game” in the context, the model’s self-attention process links up a massive web of semantic associations about how checkers games should work, and that semantic baggage throws things off.

A couple of tricks can help AI coders navigate around these limitations. First, avoid contaminating the context with irrelevant information. Second, when the agent gets stuck, try this prompt: “What information do you need that would let you implement this perfectly right now? What tools are available to you that you could use to discover that information systematically without guessing?” This forces the agent to identify (semantically link up) its own knowledge gaps, spelled out in the context window and subject to future action, instead of flailing around blindly.
— Read on arstechnica.com/information-technology/2026/01/10-things-i-learned-from-burning-myself-out-with-ai-coding-agents/