microsoft
Is is possible to support Google Agent Development Kit (ADK) in the future?
I think ADK is already supported by AgentOps which is the tracer we used. Could you try it out?
Thanks, I will check it out. The README says: Build with ANY agent framework (LangChain, OpenAI Agent SDK, AutoGen, CrewAI, ...); or even WITHOUT agent framework (Python OpenAI). You name it! 🤖 Where I didn't find ADK, so I was confused. If it is supported, could you edit the README?
We won't claim the support in the list if we haven't written an example to test it.
There is no blocking issue to support ADK. We just don't have enough people to write examples.
Hi @ultmaster! I can help to check the ADK support with the framework and write neat examples to test. Is this is priority?
Hi @ultmaster! I can help to check the ADK support with the framework and write neat examples to test. Is this is priority?
Hi thanks for your interest. I think we would very much interested in an example for Google ADK. However, depending on the size and quality of the example, we might merge it as an official example or listed as community.
So I'd like to confirm a few things first.
- Do you have any specific tasks you are trying to solve?
- What's the algorithm you will use?
- Is it mainly for agent developers or algorithm researchers?
Hi @ultmaster
I’m thinking about two example directions that highlight both enterprise and everyday-user value:
Tasks / Use Cases
1) Enterprise: Ticket Triage + Escalation Agent
- Classifies support cases
- Decides when to self-resolve vs escalate
- Learns escalation thresholds from feedback
2) Personal tool-use: Personal Productivity Task Router
- Converts natural-language tasks (e.g., “schedule call Tuesday 2pm”, “store this receipt”)
- Routes actions to Google Workspace tools (Calendar, Docs, Gmail, Sheets, Drive)
- Improves tool selection accuracy over time
Algorithm
For both examples, I will primarily use Reinforcement Learning for agent improvement, supported by:
- Outcome-based rewards (task success, escalation correctness, tool-call success)
- Optionally combining Automatic Prompt Optimization / SFT as baselines for comparison
This aligns with Agent Lightning’s philosophy of iterative, feedback-driven agent improvement vs static prompting.
Audience
Mainly agent developers building practical workflows, but structured so research users can study:
- trajectories
- reward signals
- learning curves
- baseline vs improved policy behavior
Happy to start scaffolding both examples if this aligns with the priorities!
Ticket Triage + Escalation Agent
-
Option 1 — Synthetic dataset generation: Generate realistic support tickets using an LLM to cover categories, severities, and escalation cases. This works well since models have seen similar support distributions and can bootstrap training effectively.
-
Option 2 — Public dataset: Kaggle: Multilingual Customer Support Tickets https://www.kaggle.com/datasets/tobiasbueck/multilingual-customer-support-tickets This can be used to ground the language style and supplement synthetic ticket data.
Personal Productivity Task Router
-
Option 1 — Custom dataset creation: Create a labeled set mapping natural-language user tasks to structured actions and tools (e.g., Calendar event → Gmail draft → Sheets entry). This allows fine-tuned alignment with Google Workspace actions and workflows.
-
Option 2 — Public dataset: Hugging Face: Glaive Function Calling v2 https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2 Contains natural language → function-call examples, helpful as a base for tool-routing logic.
Ticket Triage + Escalation Agent
Option 1 — Synthetic dataset generation: Generate realistic support tickets using an LLM to cover categories, severities, and escalation cases. This works well since models have seen similar support distributions and can bootstrap training effectively.
Option 2 — Public dataset: Kaggle: Multilingual Customer Support Tickets https://www.kaggle.com/datasets/tobiasbueck/multilingual-customer-support-tickets This can be used to ground the language style and supplement synthetic ticket data.
Personal Productivity Task Router
Option 1 — Custom dataset creation: Create a labeled set mapping natural-language user tasks to structured actions and tools (e.g., Calendar event → Gmail draft → Sheets entry). This allows fine-tuned alignment with Google Workspace actions and workflows.
Option 2 — Public dataset: Hugging Face: Glaive Function Calling v2 https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2 Contains natural language → function-call examples, helpful as a base for tool-routing logic.
I'm not sure I have a complete understanding or not. Would you point me to an existing ADK agent which will perform on those above-mentioned tasks. Or would you prefer writing that from scratch?
I'm not sure whether we will be able to manage the risks in that case. You will first need to verify the agent works by itself, and then prove that the agent improves after training.
