Google Agent Development Kit (ADK)

• LlmAgent = model + instruction + tools + sub-agents; the model can transfer control to a sub-agent by name (an AgentTool alternative wraps an agent as a callable tool instead).
• Workflow agents — SequentialAgent, ParallelAgent, LoopAgent — orchestrate children with code, not prompts; mixing them with LLM agents yields explicit control flow.
• Session, State, Memory: SessionService stores events and a key-value state shared across agents (output_key writes an agent's result into it); MemoryService is the cross-session store.
• Runner drives the event loop and yields typed events (text, tool call, state delta) suitable for streaming; callbacks (before_model, after_tool) act as guardrail and instrumentation hooks.
• Tools include Python functions, OpenAPI specs, MCP servers, and Google-hosted tools (Search, Code Execution); an adk web UI traces runs locally.

• Teams building on Gemini and Google Cloud who want managed deployment (Agent Engine) and Vertex tooling.
• Hierarchies combining deterministic stages with LLM agents.
• Voice/streaming agents using Gemini Live via the bidi-streaming support.

• Gemini-first: other models work through LiteLLM adapters, with weaker support for provider-specific features.
• The sub-agent transfer mechanism relies on the model choosing to delegate; misrouting shows up as silent wrong-agent answers unless you trace.
• Shared state dict is untyped; agents reading keys other agents write is a source of hard-to-find coupling.
• Young project with rapid releases; TypeScript and Java ports trail the Python API.
• Deployment story is strongest on Google infrastructure; self-hosting durable sessions means wiring your own SessionService backend.

• No Google Cloud footprint and no Gemini requirement — the ecosystem advantages disappear.
• Simple single-agent tools where the hierarchy machinery is dead weight.
• You need a mature, widely documented graph runtime today.

1from google.adk.agents import LlmAgent, SequentialAgent
2from google.adk.runners import InMemoryRunner
3 
4def fetch_metrics(service: str, window_h: int = 24) -> dict:
5    """Return error rate and p95 latency for a service."""
6    return metrics_api.query(service, window_h)
7 
8analyst = LlmAgent(
9    name="analyst", model="gemini-2.5-flash",          # model id is version-sensitive
10    instruction="Pull metrics for the named service and state anomalies as bullets.",
11    tools=[fetch_metrics], output_key="analysis",        # writes into session.state
12)
13writer = LlmAgent(
14    name="writer", model="gemini-2.5-flash",
15    instruction="Write a 3-sentence incident summary from {analysis}.",  # reads session.state
16)
17pipeline = SequentialAgent(name="incident_report", sub_agents=[analyst, writer])  # deterministic order
18 
19runner = InMemoryRunner(agent=pipeline)
20async for event in runner.run_async(user_id="u1", session_id="s1", new_message=text("checkout-api")):
21    if event.is_final_response():
22        print(event.content.parts[0].text)