Multi-agent orchestration costs more than microservices and is even slower in some cases.
Each LLM inference consumes 10–100x more energy than a SQL query. Token = CO₂. The cost is not just monetary.
Multiple agents = multiple failure points, prompt drift, prompt versioning. The system degrades silently.
A SQL bug has a stack trace. A failing agent is a black box. You can't always understand why it said what it said.