Skip to content- Sources: primary, discussion
- Summary: Aayush Garg runs GLM-5.3-Flash on a 4-GPU H200 server with 512K context and configures OpenCode so a GPT-6.1 Sol orchestrator delegates to GLM explorer and implementer subagents and verifies diffs itself. He reports that Codex and Claude Code subagents cannot target a different endpoint, so the split is not possible there. He adds ordered permission rules because OpenCode does not sandbox the agent.
- Why it matters: It shows a concrete per-agent model routing and permission config for cutting frontier model spend, plus the harness limits that block it in Codex and Claude Code.