Too Many Agents: Deciding When to Use Agents, MCP Tools, Skills, Prompt or Just a Script

As multi-agent systems become popular, many teams are beginning to turn every workflow into an agent—even when simpler approaches would be safer and easier to maintain. While building MCP-driven troubleshooting workflows for infrastructure systems, we encountered recurring cases where agent-based orchestration added unnecessary complexity, introduced retry storms, and made debugging harder than traditional automation. In several situations, replacing agents with well-designed tools or scripts significantly improved reliability and observability. This session introduces a practical decision framework for choosing between scripts, MCP tools, reusable skills, and agents. We will explore real orchestration tradeoffs, common failure patterns caused by excessive agent usage, and patterns for simplifying multi-agent systems without sacrificing flexibility. Rather than focusing on how to build more agents, this talk focuses on how to build better systems by choosing the right level of automation.