LLM Agents Interview Questions #19 - The Monolithic Agent Trap

When one ReACT loop handles both discovery and code generation, the reasoning trace becomes polluted and the model loses the signal needed to produce correct patches.

You’re in a Senior AI Engineer interview at Google DeepMind and the interviewer asks:

“Your monolithic coding agent is handling both repo-wide search and patch generation, but as the context window fills up, the patch quality tanks. How do you architect the agent loop to fix this degradation?”

Most candidates say: “Just upgrade to a model with a 1M+ token context window, use RAG to aggressively filter the search results, or prompt the agent to self-summarize.”

Wrong approach.

The reality is that more context isn’t better context. When you force a single ReACT agent to act as both a librarian (searching files) and a surgeon (writing patches), you suffer from Context Pollution. The model experiences massive Attention Dilution. Its reasoning gets hijacked by thousands of tokens of irrelevant search paths, dead-end grep results, and intermediate tool-call syntax.

Think of it like making a cardiac surgeon read the entire hospital’s medical archive while they are actively trying to perform a bypass. They don’t need the archive, they just need the final diagnosis.

Here is how you actually restructure the loop for production: