# Client-executed tool loop wiring

1. Define each tool with a `name`, a `description`, and an `input_schema` (JSON Schema). Pass the tools array on every request in the loop, not just the first one.
2. Send the request. If the response has `stop_reason` of `tool_use`, the assistant message contains one or more `tool_use` blocks. Your code executes each call.
3. Append TWO messages before the next call: first the assistant message with its content blocks exactly as returned (do not drop or reorder blocks), then a user message containing one `tool_result` block per `tool_use` block, each with the matching `tool_use_id` and your `content`. The documented shape is `{"role": "user", "content": [{"type": "tool_result", "tool_use_id": ..., "content": ...}]}`.
4. Repeat until `stop_reason` is something other than `tool_use` (usually `end_turn`). Cap the loop with a max-iterations guard; a misbehaving tool or a model that keeps calling tools is otherwise an unbounded spend.
5. Know which tools are server tools: `web_search`, `web_fetch`, `code_execution`, and `tool_search` run on Anthropic's infrastructure and you see results directly. Anthropic-defined schemas like `bash` and `text_editor` run in your application; you execute them and return `tool_result` yourself.
6. Validate tool inputs against your `input_schema` before executing. The model can emit malformed arguments; your executor is the trust boundary.

Failure modes this prevents: dropping the assistant blocks on re-send and breaking tool_use_id linkage; forgetting tools on follow-up calls so the model cannot call them again; infinite tool loops with no iteration cap; executing unvalidated input.