Add MultiAgent Semantic Conventions

[PRATIBHA-Moogi]

Area(s)

area:gen-ai

What's missing?

Current semantic conventions don't cover MultiAgentic System attributes to define MultiAgentic System space.
MultiAgent-->Tasks-->Agents-->Tools is a topological view that should cover attributes to define MultiAgent, Task, Agents, Tools specific attributes. This will help us discover MultiAgentic System topology view in a standardised manner and help us draw correlations among all key attributes to draw deeper insights on health and performance monitoring on such complex systems (or agentic workflows).

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

Based on the work on "A Taxonomy of AgentOps for Enabling Observability of
Foundation Model based Agents" we need to have coverage on all the attributes of MultiAgentic systems which govern Agenitc Behaviour & its performance in a holistic manner.
Ref Publication [https://arxiv.org/html/2411.05285v1]

Based on the taxonomy published by the above work, MultiAgentic systems can follow different topologies or workflows defined over multiple agents (a set of experts).

Task can be composed over a specific Agent call or a set of Agents, with a Gen AI model, given Task description, given expected output specification etc. It can also follow a specific topology that of defining a DAG or a workflow over multiple agents.

Here, Task key attributes such as Task description, Agents involved, expected_output can be added to the semantic conventions to well define Task resources.

Each Agent further can be composed with a bundle of tools required to carry out Agent's goal given its role, given a prompt_template, in-context info. So Agent Goal, Role, Task, Tool, Expectated output can characterise individual Agents attributes and can be added to the semantic conventions for covering Agent resources.

Each tool also can be characterised with a set of attributes - tool_type, tool_description, tool_argument etc. So those attributes can also be added into semantic conventions.

Each prompt can also be described by means of prompt_template_type, prompt_template_info. And further to that other constructs of prompt_template_info - User Goal, Instruction, Query, Few-Shot examples, Output format ask, Tools List, ChatHistory etc can also be added as a set of attributes to characterise Prompts.

The above some of the key MultiAgentic System attributes, if added to the semantic conventions, can further enable rich set of Observability views as the below ;

1. Discovery of MultiAgentic system topology view
2. Discovery of Tasks, Task-Agent relationship
3. Discovery of Agent-Tool relationship
4. Discovery of Agent-Prompt attributes relationship
5. Discovery of Tools usage, Tools Types & associated performance metrics
6. Correlations of Agent's performance with Prompts attributes
7. System level metrics associated at each components of MultiAgentic System

### ##Describe the solution you'd like

A new semantic conventions with the key attributes to describe multiagentic systems / applications.

/cc @gyliu513

Describe the solution you'd like

No response

[gyliu513]

Related with #1530

[TaoChenOSU]

We should distinguish between single agent and multi-agent workflows.

We may only need a very thin layer for multi-agent workflows in the GenAI conventions. It's because the underlying mechanism of multi-agent-workflows is message passing, which is not GenAI specific.

The things we should probably trace may include the following:

• Attributes
  • The agents in a team
  • The pattern employed by the team (Round robin, group chat style, MagenticOne, etc.)
• Events
  • Start event
  • Transfer events (from one agent to another)
  • Termination event

[gyliu513]

@TaoChenOSU totally agree with you, besides this, I think we also need to consider introducing some new attributes, I think there should be at least two new concepts for mulit agent, including flow/workflow and task.

I was working with @PRATIBHA-Moogi for this, but glad to discuss here, thanks!