Great Part 3 — the hands-on flow clarifies what was abstract in earlier parts. One thing I've learned building agent systems with Claude plus MCP for clients: treat the tool interface like a product surface, not an implementation detail. Clear tool descriptions and tight input schemas cut hallucinated calls drastically. Curious how ADK handles tool-selection once you add 10-plus tools — does context bloat become an issue, or is there a retrieval layer in the SDK.