That's why we built AgentPrism Workflows (OSS, Apache-2.0).
Code-defined workflows - agent() / parallel() / pipeline() in plain JS - where every agent() call picks its own harness AND model:
• Fable for design judgment
• gpt-5.6 (in whichever harness it's best in) for implementation
• GLM via OpenCode for the mechanical sweeps
• any custom ACP agent for the weird stuff
It drives the real Claude Code / Codex / OpenCode over ACP - their tool loops, their permissions, your existing auth. No reimplemented agent loop.