adk-demo

A minimal demo of deploying a Google Agent Development Kit (ADK) agent to LangSmith Agent Server, following the Deploy Google ADK agents guide.

The agent is a personal task manager: a single ADK LlmAgent whose tools read and write a per-session to-do list via tool_context.state. Because state is persisted through saf_sdk.adk.wrap and LangsmithSessionService, tasks survive across turns (and restarts) on the same thread.

How it works

agent.py defines:

• Tools that mutate a task list stored in session state:
  • add_task(title, priority) — add a task (low / medium / high)
  • list_tasks(include_done) — list open (or all) tasks, sorted by priority
  • complete_task(task_ref) — mark a task done by id or exact title
  • delete_task(task_ref) — remove a task by id or exact title
  • summarize() — counts by status and priority
• An Agent (task_manager) using gemini-2.5-flash that drives those tools.
• A module-level agent produced by wrap(Runner(...)), which Agent Server imports and serves.

Two things make this deployable:

1. The runner uses LangsmithSessionService() so session state is stored in the deployment's checkpoint store.
2. The wrapped graph is exported as the module-level agent symbol, which langgraph.json registers under the graph id adk_echo.

Project layout

adk-demo/
├── agent.py          # exports the wrapped ADK agent
├── langgraph.json    # Agent Server config (graph id: adk_echo)
├── pyproject.toml    # Python dependencies
├── uv.lock           # pinned dependency lockfile
├── .python-version   # pinned Python version (3.12)
├── .env.example      # environment variable template
└── README.md

Prerequisites

• uv (package manager; installs Python 3.12 for you via .python-version)
• A LangSmith API key
• A Gemini API key (the agent uses a Gemini model)

The LangGraph CLI is pulled in as a project dependency (langgraph-cli[inmem]), so syncing the project gives you the langgraph command.

Setup

Install dependencies (creates a virtualenv and installs from the lockfile):

uv sync

Copy the environment template and fill in your credentials:

cp .env.example .env

GEMINI_API_KEY=your-gemini-api-key
LANGSMITH_API_KEY=your-langsmith-api-key
LANGSMITH_TRACING=true
LANGSMITH_PROJECT_NAME=adk-demo

Run locally

uv run langgraph dev

This serves the agent at http://127.0.0.1:2024 and opens LangSmith Studio. Try messages like:

• "Add 'write the demo readme' as a high priority task"
• "What's on my list?"
• "Mark the readme task done"
• "How am I doing?"

Or call it over HTTP:

# Create a thread
THREAD=$(curl -s -X POST http://127.0.0.1:2024/threads \
  -H "Content-Type: application/json" -d '{}' \
  | python -c "import sys, json; print(json.load(sys.stdin)['thread_id'])")

# Run the agent and wait for the final response
curl -s -X POST "http://127.0.0.1:2024/threads/$THREAD/runs/wait" \
  -H "Content-Type: application/json" \
  -d '{
    "assistant_id": "adk_echo",
    "input": {"messages": [{"type": "human", "content": "Add buy milk as a low priority task"}]}
  }'

Reusing the same THREAD across requests demonstrates that the task list persists in session state.

Deploy to LangSmith

uv run langgraph deploy --name adk-demo

See Deploy to cloud for environment configuration, deployment types, and revision management.