Gemma 4 31B hits endless loops in coding agents

Local agentic coding remains a brittle frontier where even highly capable models like Gemma 4 can fail due to systemic prompt mismatch.

• –The 31B model's reasoning mode can get stuck thinking for over 10 minutes without making a tool call, eventually devolving into endless garbage token generation.
• –In agents like Cline, the model frequently repeats clarification requests or fails to transition from reading files to executing edits.
• –The issue highlights a broader challenge in local AI: raw model capability doesn't automatically translate to reliable agentic tool-use without specialized system prompts.
• –Many developers running local GPUs are finding Qwen-based models to be more reliable out-of-the-box alternatives for autonomous coding workflows.