Why it matters

ADK is Google's answer to LangChain and CrewAI. Understanding its primitives shapes how you build agents on Google infrastructure.

Advertisement

The architecture

Core: Agent (top entity), Tool (callable functions), Runner (orchestration), Session (per-conversation state), Memory (cross-session facts).

Agent types: LlmAgent (LLM-driven), workflow agents (sequential, parallel, loop), custom.

ADK Python primitivesAgentLlm or workflowTools + Sessioncallables + stateRunnerorchestratesSame conceptual model as ADK Java; Python-first with Gemini integration
ADK core.
Advertisement

How it works end to end

Model abstraction: pluggable model layer. Gemini native; other providers via adapters.

Session: SessionService for persistence. In-memory for dev, Vertex-backed for production.

Deployment: local, Cloud Run, Vertex AI Agent Engine.

ADK Web: bundled UI for testing agents interactively.