Google ADK Python agents

Guides code-first Google ADK agents, multi-agent systems, A2A, MCP tools, workflow agents, state, memory, deployment, and evaluation.

Identify use case

Choose structure

Select agent type

Wire tools

Rule 01

ADK agents should be code-first so behavior is versioned and testable.

Rule 02

Use workflow agents for predictable control flow and LlmAgent when dynamic reasoning is needed.

Rule 03

State, memory, artifacts, callbacks, plugins, and observability are deliberate architecture choices.

Rule 04

Evaluation with adk eval belongs before deployment, not after incidents.

Core lanes from input to output, aligned with the command's real execution path.

Start

Input, scope, route

1 Identify use case Decide whether the work is a single agent, multi-agent system, A2A integration, MCP toolset, workflow, state, memory, callback, deployment, or evaluation.
2 Choose structure Create the package convention with __init__.py and agent.py exposing root_agent or app.

Work

Agent / skill execution

3 Select agent type Use Agent or LlmAgent for dynamic reasoning, SequentialAgent for fixed pipelines, ParallelAgent for concurrency, LoopAgent for iteration, or RemoteA2aAgent for remote agents.
4 Wire tools Add Python tools, MCPToolset, LongRunningFunctionTool for human-in-loop, sub_agents, callbacks, plugins, or artifacts as needed.

Verify

Gate, review, validation

5 Manage state Use ToolContext.state for ephemeral data, MemoryService for long-term memory, and artifacts for saved outputs.
6 Run locally Use adk web for dev UI, adk run for CLI testing, or adk api_server for a production FastAPI server.

Report, handoff, artifact

7 Evaluate Run adk eval with evalset JSON before deployment.
8 Deploy Choose Cloud Run, Vertex AI Agent Engine, or GKE using the deployment reference.

Production App pattern

Use App when plugins, event compaction, artifacts, or custom lifecycle management are needed.

Workflow agents

Sequential, parallel, and loop agents make predictable pipelines explicit.

Concrete invocations for each flag, subcommand, mode, or route available in the live workflow.

Multi-agent system Recommended

/ak:google-adk-python multi-agent support bot with MCP tools

Use when:: You need ADK architecture for specialized sub-agents and tools.
Expected:: Chooses agent types, package convention, MCPToolset wiring, state handling, and evaluation path for the requested app.

A2A integration

/ak:google-adk-python A2A remote agent integration

Use when:: Remote agents should communicate through A2A protocol.
Expected:: Loads the multi-agent and A2A reference and outlines RemoteA2aAgent usage for cross-agent coordination.

Vertex deployment

/ak:google-adk-python deploy agent to Vertex AI Agent Engine

Use when:: An ADK agent is ready for managed deployment.
Expected:: Uses the deployment reference and includes evaluation checkpoints before recommending rollout to production.

Next