AI Agent Semantic Convention

[gyliu513]

Area(s)

area:gen-ai

Is your change request related to a problem? Please describe.

Based on the white paper from Google from google https://www.kaggle.com/whitepaper-agents , we need semantic convention for a SINGLE AI Agent, the SINGLE AI Agent will be composed by LLM and Tools, and Tools will contain DataStore, Functions, Extensions

/cc @lmolkova

Describe the solution you'd like

A new semantic convention for ai agent

Describe alternatives you've considered

No response

Additional context

No response

[TaoChenOSU]

Thank @gyliu513 for creating this issue to continue our discussion from the WG call.

When we invoke an agent to perform a task, the agent invokes the LLM one or more times and can potentially run tools that are available to it. We could reuse the existing gen_ai convention for the LLM calls. I am not sure if a convention for tool will be necessary since tools can take on different forms, such as a code executor, an external service call, and data retrieval, etc. Each of the type should have its own conventions instead and some of them already exist. For example, data retrieval should follow the (vector) database convention.

I do agree we need to capture information about an agent, including id, name, instruction, and different settings.

[joaopgrassi]

@gyliu513 @TaoChenOSU any of you will be taking this? Given it is already part of the GenAI project, I removed the "needs triage" label but would be good to get it assigned to someone from the SIG.

[gyliu513]

@joaopgrassi I was working for the PR, hope I can have a draft next week, thanks