The @ollama cloud subscription gives broad access to open-weight models, and usage vs. credits is fairly generous. However, if you run kimi-k3 or glm-5.3 as the main model, the allocation burns through fast. @goose_oss is the harness that fits this plan best. Keep the main agent on deepseek-v4.1-flash and define subagents as goose recipes. A recipe is the subagent config. Lock role, extensions, and model into the recipe, and the subagent spins up with that exact setup when invoked. Role split: • Hard, long-horizon reasoning / difficult coding: glm-5.3 • Document and external-repo analysis: glm-5.3-flash • Web search: a cheap recipe with search tools attached (deepseek-v4.1-flash or glm-5.3-flash) Put into .goosehints when to call which recipe and when the main agent should handle the work itself. As the task proceeds, the main agent only invokes the recipes it needs. Expensive models get used only for those segments, so you save tokens. Other harnesses often support per-subagent models too, but wiring Ollama Cloud’s open-weight models into reusable recipes is where goose fits best.