---
description: Google Agent Development Kit rules for agents, tools, sessions, memory, artifacts, evaluation, and deployment
globs: ["**/*.py", "**/*.ts", "**/*.tsx", "**/*.go", "**/*.java", "pyproject.toml", "package.json"]
alwaysApply: false
---

# Google ADK Rules

## Agent Design

- Keep each agent focused on a clear goal, persona, and tool set.
- Use LLM agents for flexible reasoning and workflow agents for deterministic orchestration.
- Write instructions that define task boundaries, tool-use rules, and escalation behavior.
- Split multi-agent systems by responsibility rather than by implementation convenience.
- Keep model choices configurable.

## Tools

- Give tools narrow, typed inputs and outputs.
- Validate tool arguments before performing side effects.
- Keep secrets, credentials, and privileged APIs out of agent prompts.
- Handle tool errors explicitly and return actionable failure messages.
- Be aware of ADK tool limitations; some built-in tools cannot be combined with other tools on the same agent.

## Sessions, State, and Memory

- Use session state for current-conversation data.
- Use memory for cross-session recall and retrieval.
- Keep state small and serializable.
- Do not store large files or binary payloads in session state.
- Make state keys stable and documented.

## Artifacts

- Use artifacts for generated files, uploaded files, reports, images, audio, and other binary data.
- Configure an artifact service in the runner before relying on artifact operations.
- Version artifact filenames intentionally and avoid overwriting semantically different outputs.
- Store only references or summaries in state when full content belongs in artifacts.

## Evaluation and Deployment

- Add tests for tool behavior, agent routing, prompt regressions, and unsafe tool calls.
- Use trace or event logs to debug agent decisions.
- Keep local development, staging, and production configuration separate.
- Add observability for latency, tool failures, token use, and handoff failures.

## Common Mistakes

- Do not make one agent responsible for every workflow.
- Do not let tools accept arbitrary shell, SQL, or HTTP input without validation.
- Do not rely on prompt text for access control.
- Do not hide important side effects behind generic tool names.