The cleanest way is the `Agent` tool built into Claude Code. You describe the task in the `prompt` parameter and a subagent picks it up directly -- no middleware layer. ```python Agent({ description: "Brief label for what this subagent does", subagent_type: "general-purpose", # or "Explore", "Plan", etc. prompt: "Detailed, self-contained instructions. Include all context the agent needs -- it has no memory of this conversation." }) ``` A few things to keep in mind for clean handoffs: - **Self-contained prompts**: The subagent starts cold. Give it everything: file paths, what you've already tried, what decision to make, what format to return results in. - **Run in background** when results aren't needed immediately: `run_in_background: true` lets you continue working in parallel. - **Foreground by default** when you need the result before your next step. For external/remote triggers (e.g. kicking off an agent from a webhook or CI), the `RemoteTrigger` tool handles that -- but that requires loading its schema first via `ToolSearch`. Is your question about Claude Code's Agent SDK specifically, or about a broader architecture where two separate systems need to exchange tasks directly? The answer differs depending on whether you're orchestrating within a single Claude Code session or across services.