Use this before shipping a small AI agent, MCP server, browser agent, support workflow, or tool-calling demo to real users.
- List every tool the agent can call.
- Mark tools that read files, write files, send messages, open URLs, run commands, call APIs, update memory, access CRM/support data, or touch customer-visible state.
- Separate read-only tools from tools that change state.
- Remove any tool the first launch does not need.