PROJECT THREE - TOOLS, THE MODEL CONTEXT PROTOCOL (MCP), AND SAFETY

LEARNING GOALS By the end of this chapter you will be able to connect your harness, as a client, to a separate tool server using the MCP. You will be able to add hooks, including a human approval step before risky actions. You will be able to find and fix an idempotency bug. You will be able to defend the agent against a prompt injection and a poisoned tool description. WHY TOOLS AND SAFETY ARRIVE TOGETHER Think about handing a new employee a ring of keys on the first day. Each key makes the employee more useful, and each key also makes a mistake more costly. Tools are the keys of an agent. So every time we add a tool, we must also ask who might misuse it, and what could go wrong if the agent is tricked. This chapter adds tools and safeguards in one move.