OpenCode users push multi-GPU agent routing

This is less a one-off setup question than a glimpse at the next bottleneck for local AI coding stacks: orchestration, not raw tokens per second.

• –OpenCode explicitly supports custom `llama.cpp` providers via `@ai-sdk/openai-compatible`, so separate local endpoints can be wired in today as distinct providers with different `baseURL` values.
• –Its agent system also allows per-agent model overrides, which means advanced users can manually shard work across models or endpoints instead of relying on one global default.
• –The gap is subagent routing: OpenCode’s own docs say subagents inherit the parent model unless explicitly configured, and community feature requests are still asking for dynamic model selection on subagent tasks.
• –Another active issue asks for multiple instances of the same provider, which is closely related to this Reddit post’s goal of treating several local inference backends as first-class resources instead of hacks.
• –A broader “agent teams” design proposal in OpenCode’s GitHub points toward parallel, multi-model teammates, which is exactly the direction local multi-GPU operators want for serious self-hosted coding workflows.