GSoC 2026 — Project #18: Prior work and scope questions (Agentic Toolkit)

[LuciferDono]

Hi @freddychiu @ravi9,

I'm interested in Project #18. I've been working with OpenVINO and agentic frameworks recently. Two relevant PRs on openvino_notebooks:

• PR #3330: Self-correcting RAG agent using LangGraph + OpenVINO, with document grading, query rewriting, and conditional routing via state machine
• PR #3332: Moondream2 VLM notebook with INT4 compression and Gradio demo

The RAG agent in #3330 already does some of what the toolkit would need: wrapping an OpenVINO model into a LangGraph tool chain, managing retrieval context locally, and routing based on agent state. Extending that into a reusable toolkit with proper APIs is a natural next step.

A few questions on scope:

1. For the agent creation APIs, should the toolkit expose its own abstraction layer, or provide integration adapters for existing frameworks (LangChain's BaseLLM, LlamaIndex's CustomLLM)? The second approach gets adoption faster but couples the toolkit to external APIs.

2. The description mentions supporting both C/C++ and Python. For the GSoC scope, is it reasonable to focus on Python first with the C++ bindings as a stretch goal?

3. For the A2A (agent-to-agent) deliverable, is there a specific protocol in mind (Google's A2A protocol, or something custom)?

Happy to start prototyping once I have a sense of direction on these.

[ravi9]

HI @LuciferDono , Apologies for the delayed response !

1. I think the best long-term structure is a lightweight toolkit-owned abstraction layer with framework adapters layered on top. Some of the existing workflows:

• llama_index_agent
• openai_agent.py
• langchain-openvino

I also added some more of my thoughts here

2. Yes, focusing on Python first sounds good.

3. Sure, we can start with Googles A2A

Please consider submitting your proposal formally through the GSoC portal !

Thanks,
Ravi
cc: @freddychiu

[LuciferDono]

Thanks Ravi, appreciate the links. I'll go through the model_server demos and langchain-openvino this week.

For the abstraction layer, should it sit on top of model_server's gRPC/REST API, or wrap openvino-genai directly as a standalone Python lib?

Working on the proposal now, aiming to have it on the portal by end of this week.

[ravi9]

Hi @LuciferDono , Sure.
Yes, the abstraction layer should be on top of OVMS (model_server).