Databricks co-founder Matei Zaharia says choosing the right agent harness can halve LLM-based coding cost

Task-aware routers and model harnesses are going to be the next big thing. This is the only way to keep costs in check while extracting maximum performance. I have been experimenting with this in security, specifically tasks like malware analysis, where MiniMax M3 is almost comparable to Opus 4.8. But if you take another task like threat intel reasoning, GLM-5.2 is better suited and provides performance comparable to Opus 4.8.