Google Agent Development Kit (ADK)

This skill covers building production-grade AI agents with Google's Agent Development Kit (ADK) for Python, including agent composition, tool design, session/state/memory management, artifacts, evaluation, and deployment.

Workflow for Building an ADK Agent

1. Define the agent — Create an LlmAgent (or Agent) with a clear name, model, instruction, and description. Keep the instruction focused on one job.
2. Author tools — Write plain Python functions with type hints and docstrings, or wrap existing APIs with FunctionTool. Validate all inputs before side effects.
3. Compose multi-agent systems — For complex workflows, split responsibility across sub-agents and use SequentialAgent, ParallelAgent, or LoopAgent for deterministic orchestration, or delegate via sub_agents for LLM-driven routing.
4. Wire up session and state — Choose a SessionService (in-memory for dev, DatabaseSessionService or Vertex AI-managed for production) and use session.state for conversation-scoped data.
5. Add memory (optional) — Configure a MemoryService for cross-session recall when the agent needs to remember facts between separate conversations.
6. Handle artifacts (optional) — Configure an ArtifactService when the agent generates or receives files, images, or other binary outputs.
7. Run locally — Use adk web, adk run, or the Runner API to exercise the agent against a Session.
8. Evaluate — Write .evalset.json test cases and run adk eval to check tool-call trajectories and response quality against regressions.
9. Deploy — Package the agent for Vertex AI Agent Engine, Cloud Run, or GKE, and separate dev/staging/prod configuration.

Agent Design