Build a Multi-Tool Agent with Escalation Logic

Why it matters

Section 1.1 decides who drives the loop and when it stops. Here you build that loop and wrap it in tool contracts, error metadata and a policy gate, so each decision meets real tool calls.

What to practise

Write three or four MCP tools whose descriptions set out purpose, inputs and boundaries, two of them near-identical except for their descriptions. Drive the loop from stop_reason: run tools on tool_use, answer on end_turn. Return errors with an errorCategory (transient, validation or permission), an isRetryable flag and a readable explanation; the agent should retry transient errors and explain the rest. Add a hook that intercepts tool calls, blocks operations above a threshold amount and redirects to escalation. Then send a message raising several issues and check that one unified answer handles each.

How it fits

The loop is 1.1.1 and 1.1.2; compare your stop logic with the anti-patterns in 1.1.4. In 1.4, the hook is the policy gate of 1.4.2, its escalation should carry what 1.4.4 lists, and the multi-issue test is 1.4.3. In 1.5, the block must hold whatever the prompt says (1.5.1, 1.5.3). In 2.1, the near-identical tools exercise 2.1.1 and 2.1.2. In 2.2, the error metadata is 2.2.2, and a rejection should read like 2.2.3.