What problem does it solve? Multi-agent designs that concatenate the coordinator's full history into a single agent dilute model attention, leak policies across tasks, inflate cost by running everything on an expensive model, and widen the blast radius of prompt injection. This kata teaches structural isolation so each subagent starts with an empty context, its own tools, and its own model. ## Core Features & Use Cases - Structural subagent isolation: Register subagents as AgentDefinition entries in ClaudeAgentOptions.agents and dispatch them via the built-in Agent tool, so every dispatch opens a fresh session with its own system_prompt, tools, and model. - Per-subagent model and tool assignment: Run cheap extraction on haiku with tools=[] while reserving the expensive model for coordinator synthesis, and receive only each subagent's final message as a tool_result. - Use Case: In a Multi-Agent Research scenario over 30 documents, dispatch one isolated extractor subagent per document so a poisoned document contaminates only its own session while the coordinator aggregates just the extracted facts. ## Quick Start Ask the assistant to redesign your multi-agent research pipeline so each document is processed by an isolated extractor subagent registered with AgentDefinition and dispatched via the Agent tool.
