OpenCode users tune hybrid local agent stacks

This is less a product announcement than a snapshot of where AI coding workflows are heading: one orchestrator up top, several specialized local models underneath, and constant tradeoffs between cost, latency, and autonomy.

• –It highlights why OpenCode’s provider-agnostic design matters: users can mix API and local models instead of locking the whole stack to one vendor.
• –The most interesting choice is using a non-coder model for review, which reflects a growing belief that diversity across agents can catch failure modes a single coding model misses.
• –Community feedback focused on practical tuning, not hype: swap GLM-4.5-Air into build, test smaller quants for better throughput, and push more hierarchical planning local.
• –The real bottleneck sounds less like benchmark scores and more like long-run agent drift, which is exactly the kind of systems problem local-first power users are now optimizing around.