Google's Kotlin agent SDK skips reflection for compile-time tools

Google announced ADK for Kotlin 1.0 on September 9, and the change that matters is how tools get wired up. Most agent SDKs, including the Python ADK, resolve tools at runtime through reflection, inspecting function signatures as the program runs. ADK for Kotlin declares tools through KSP at compile time instead. There's no runtime reflection step to survive.

That's a direct fix for two things mobile apps care about. R8 code shrinking strips anything it can't prove is reachable, and reflection confuses it. Reflection scanning also adds to cold-start time, the delay before an app is usable after launch.

On feature parity, ADK for Kotlin matches ADK 1.0 Core: hierarchical delegation, summarization-based context management, human-in-the-loop confirmation gates, and session persistence with resumability.

The rest of the stack follows the same on-device logic. Conversation history persists to Room, memory indexes through AppSearch, and inference can run on-device via LiteRT-LM or ML Kit.

Google is explicit that this is a separate runtime from the Python ADK, not a port. Building two SDKs with different tool-binding architectures for the same agent concepts is a real engineering cost, and Google paid it anyway.

What isn't in the announcement: benchmark numbers for cold-start time or shrunk APK size against a reflection-based baseline. If the compile-time approach is the fix Google says it is, those numbers should show up once third-party apps adopt it.