New example: Google ADK by KellySit · Pull Request #309 · microsoft/agent-lightning

KellySit · 2025-11-13T15:52:30Z

New Example: ADK Agent

Adds a new example demonstrating ADK agent training with Agent-Lightning. ADK is an observability framework for agent traces with span visualization and Cloud Trace integration.

Files Added:

examples/google_adk/adk_agent.py - LitAdkAgent implementation
examples/google_adk/train_adk.py - Training script with CI support
examples/google_adk/adk_debug.py - Sanity check script
examples/google_adk/prepare_dataset.py - Dataset conversion utility
docs/how-to/train-adk-agent.md - Comprehensive training guide
Updated pyproject.toml with adk optional dependency group

…d dependencies at pyproject.toml. Created google_adk under examples (adk_agent.py, adk_debug.py, prepare_dataset.py, README.md, train_adk.py)

KellySit · 2025-11-13T15:54:28Z

@microsoft-github-policy-service agree [company="KellySit"]

KellySit · 2025-11-13T15:55:21Z

@microsoft-github-policy-service agree

…gitignore

ultmaster

Thank you for your contribution! Let's work together to polish this work!

ultmaster · 2025-11-14T05:38:46Z

@@ -0,0 +1,395 @@
+# Train ADK Agent with Agent-lightning and VERL


Let's refine this example. Users might have already read the how-to with verl + langchain. This how-to should focus on the differences -- how to make ADK agent available for training. Omitting other unnecessary details.

Thank you so much! I have updated and made a latest commit based on the feedback. Ready for the review.

ultmaster · 2025-11-14T05:39:09Z

+```
+
+This allows you to verify that the agent logic and reward computation behave as expected before reinforcement learning is introduced.
+


If that's a how-to, please write the training results here.

ultmaster · 2025-11-14T05:39:15Z

        algorithm=algo,
        # Increase the number of runners to run more rollouts in parallel
-        n_runners=8,
+        strategy={  


why this change?

ultmaster · 2025-11-14T05:39:40Z

@@ -0,0 +1,8 @@
+vendor/


I think most of these ignores have been included in global gitignore. Please check,

ultmaster · 2025-11-14T05:39:59Z

+logger = logging.getLogger(__name__)
+
+try:  # pragma: no cover - import guarded for optional dependency.
+    from google.genai import types as genai_types


no need for try except here. This example assumes user have installed google ADK.

ultmaster · 2025-11-14T05:40:29Z

+console = Console()
+
+
+def parse_args() -> argparse.Namespace:


let's work together to add CI for this example.

ultmaster · 2025-11-14T05:41:17Z

+    print(f"Wrote {len(df)} rows to {out_path}")
+
+
+def create_toy_dataset(n_train: int = 8, n_test: int = 2) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:


So you use a toy dataset. Any real dataset we can use?

ultmaster · 2025-11-14T05:41:54Z

+
+This example demonstrates training a Google ADK (Application Development Kit) agent using Agent-Lightning with reinforcement learning. The agent processes user instructions and generates ADK actions for Google applications. It's compatible with Agent-lightning v0.2 or later.
+
+## Requirements


If you have decided to create a how-to tutorial, this README can be greatly simplified and only keeps necessarily installation and running guidance.

ultmaster · 2025-11-14T05:42:11Z

-agentlightning/dashboard/**/*.js
-agentlightning/dashboard/**/*.html
-agentlightning/dashboard/**/*.svg
+.claude/*.local.json


no need to change this I think.

…nt-lightning into examples/google_adk

KellySit · 2025-11-19T10:37:16Z

@ultmaster Thank you so much for the feedback. I have made some updates based on the feedback:

Fixed unnecessary changes (examples/apo/room_selector_apo.py)
Updated adk_agent.py and adk_debug.py
Added training result showcase to docs/how-to/train-adk-agent.md
Added CI and renewed prepare_dataset.py
Omitted unnecessary things (docs/how-to/train-adk-agent.md, examples/google_adk/README.md)

The changes are now pushed and ready for re-review. The latest commit is the ready one. Thanks.

JasonHonKL · 2025-11-19T17:32:29Z

@KellySit Great work but seems you have to resolve the conflict before the CI can run !

KellySit · 2025-11-19T18:37:55Z

@ultmaster PTAL, conflicts resolved. Thank you!

ultmaster · 2025-11-20T01:44:53Z

Hi. Appreciate the improvements.

Given the implementation of current LitAdkAgent, I recommend using the style of @agl.rollout, which would be more concise. You only need to keep the async version. No need for both.

Another question is, would you mind changing a dataset? I don't want all the examples focusing on a single spider dataset. If you want to stick to spider, you can augment the existing how to train sql how-to document by adding a section about ADK; and merge the examples into spider directory. There is no need for another example directory --- which could cause confusion on why this is not called spider, do we prefer langchain over ADK, or whatever.

Also, would you allow me to edit the PR so that I can make some minor improvements and verify with CI?

ultmaster · 2025-11-20T02:21:59Z

+        instruction = self._compose_instruction(task)
+        agent_name = task["app_id"] or "adk_agent"
+
+        return LlmAgent(


I'm not sure about how ADK works. Is this agent equivalent to a single non-agentic LLM?

Basically yea, ADK agent example here wraps in its own orchestration/runtime. No use of tools or multi-step planning, so behavior is close to a non-agentic LLM, with a structured system prompt. But ADK supports multi‑step, agentic workflow too. I try to keep things simple for this example, but if necessary, we can extend it too.

@ultmaster Thanks! I will update this later, adopt the style @agl.rollout and switch to a different dataset. Feel free to edit the PR, improve it, or help with verification, let’s make this project great together!

@KellySit maybe we can discuss more on what tasks, datasets and the expected outcomes of this example. Are you available on Discord or Wechat or X?

@ultmaster Sure! I am available on Discord, here is my username: ivylavender.

@KellySit Sorry I've been laggy on responses.

It seems I can't add you on Discord:

Would you be able to DM me in our discord community: https://discord.gg/RYk7CdvDR7

johnson7788 · 2025-12-04T03:18:01Z

Great example, I love you

New example: Google ADK. Added docs/how-to/train-adk-agent.md. Update…

b326bb7

…d dependencies at pyproject.toml. Created google_adk under examples (adk_agent.py, adk_debug.py, prepare_dataset.py, README.md, train_adk.py)

KellySit added 3 commits November 14, 2025 00:27

Updated examples/adk_agent.py

ab76f1f

Updated examples/google_adk/adk_agent.py. Added examples/google_adk/.…

18721ea

…gitignore

Merge branch 'main' into examples/google_adk

84bdcdf

ultmaster reviewed Nov 14, 2025

View reviewed changes

KellySit added 2 commits November 19, 2025 17:11

Updates on Google ADK based on feedback.

f8f4c9f

Merge branch 'examples/google_adk' of https://github.com/KellySit/age…

4ddfe51

…nt-lightning into examples/google_adk

KellySit force-pushed the examples/google_adk branch 2 times, most recently from 6a05b16 to 4ddfe51 Compare November 19, 2025 10:26

Updates on Google ADK

267ac1f

Merge upstream/main and regenerate uv.lock to resolve previous conflicts

3b3f55f

ultmaster reviewed Nov 20, 2025

View reviewed changes

		@@ -0,0 +1,395 @@
		# Train ADK Agent with Agent-lightning and VERL

		```

		This allows you to verify that the agent logic and reward computation behave as expected before reinforcement learning is introduced.

		print(f"Wrote {len(df)} rows to {out_path}")


		def create_toy_dataset(n_train: int = 8, n_test: int = 2) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:


		This example demonstrates training a Google ADK (Application Development Kit) agent using Agent-Lightning with reinforcement learning. The agent processes user instructions and generates ADK actions for Google applications. It's compatible with Agent-lightning v0.2 or later.

		## Requirements

Conversation

KellySit commented Nov 13, 2025

New Example: ADK Agent

Uh oh!

KellySit commented Nov 13, 2025

Uh oh!

KellySit commented Nov 13, 2025

Uh oh!

ultmaster left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

KellySit commented Nov 19, 2025

Uh oh!

JasonHonKL commented Nov 19, 2025

Uh oh!

KellySit commented Nov 19, 2025

Uh oh!

ultmaster commented Nov 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

johnson7788 commented Dec 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

ultmaster commented Nov 20, 2025 •

edited

Loading