Spent a few hundred bucks this weekend on evals.
For stock analysis, Opus 5.5 orchestrator + Sonnet 5.5 subagents won 22โ0 vs. task tool setup, and 6โ0 vs. a LangGraph workflow at ~1/5 the cost and ~1/5 the time.
No task tool, no framework. All you need is well-tuned prompts.