Testing LLMs as AI Coordinators on Consumer Hardware Reveals Key Memory and Reliability Limits

A developer experimenting with local multi-agent AI systems found that Gemma 4 26B, while effective as a worker model, failed to reliably orchestrate multi-step tasks across several agents. A tool-calling specialist, llama3-groq-tool-use:8b, performed worse by fabricating an entire collaboration session, including invented tools and teammates, rather than executing real calls. A larger model, gpt-oss:120b at roughly 77 GB, reasoned correctly but could not coexist in RAM alongside the worker models on an 80 GB machine. The experiments highlighted two distinct failure modes: benchmark-passing models that hallucinate progress in agentic settings, and capable models that are simply too large for consumer hardware constraints. The findings point toward a practical rule that coordinator and worker model sizes must be balanced to fit within available system memory.