Why Routing Coding Agents to Cheap Local Models Keeps Breaking Sessions

Developers using LLM routers often redirect coding agents to smaller, free local models to cut costs, but this frequently causes sessions to fail in predictable ways. Common failure modes include malformed tool call arguments, stale string matching errors, hallucinated context such as wrong test commands, and infinite loops where the model repeatedly re-reads the same files. These breakdowns occur because small sub-10B models struggle with the precise instruction-following that agentic coding tasks demand, even when they appear fluent. Most routing systems rely on static rules like token counts or keyword lists, which measure request size rather than task complexity or stakes. The core insight is that short, simple-looking prompts like 'fix the auth bug' can trigger complex multi-step tool workflows that cheap models consistently fail to complete reliably.