Why MCP servers should expose fewer tools per task stage to keep agents safe

The article argues that adding tools to an MCP server is not cost-free, since each tool's name, description, and schema compete for limited context and create extra decisions for the agent. Exposing an entire tool catalog often causes predictable failures, such as read tasks invoking write tools or redundant calls to overlapping tools. Instead, the author proposes a stage-based tool budget that limits what an agent can see at each step, such as observing, preparing, executing, and verifying a deployment. Under this approach, write actions arrive late, narrowly scoped, and paired with a verification path, and destructive or overly broad operations are excluded. The piece also recommends writing a one-line usage contract for each visible tool to reduce ambiguity between overlapping tool names.