Why Local LLMs Excel at Code—But Fail at Agent Tasks

Testing six local LLMs reveals a stark gap: models scoring 90%+ on code benchmarks plummet to 17-50% on agent tasks like tool selection and chaining. The problem isn't purely size—it's architecture; a 1.8B parameter model outperformed a 7.7GB one. For developers considering local AI agents, this means today's open-weight models lack the architectural foundation for reliable autonomous tool use.