Hey everyone!
I’m currently building a 3-tier ‘Planner-Worker-Judge’ multi-agent orchestration shell (leaning heavily on Gemini/Genkit). For those of you building agentic workflows, how are you handling your ‘Judge’ or auditor tier to keep latency and costs down?
Do you prefer routing the evaluation step to a faster/smaller model (like Flash), or do you strictly use Pro models for the final safety checks?
Curious to hear your stack preferences!