Agent as a service / Team of agents as a service

[cschneider]

As multiple teams work on specialized AI agents and agent teams, there's a risk of conflicts when editing shared configurations and code concurrently. This can slow development and create inefficiencies.

Proposed Solution:

I propose extending Microsoft AI Autogen to allow teams to define agents and agent teams as independent services, giving them full control over their configurations and code. These services would communicate with each other through defined APIs or protocols, ensuring smooth interactions without interfering with each other’s work.

Key Benefits:

Reduced Conflict: Teams can work independently without impacting each other.
Autonomy: Full control over agent configurations and versions.
Efficiency: Parallel development without bottlenecks.
Scalability: Modular, flexible agent services that can evolve independently.

Does that make sense? Has anyone else already brought that up (I did not find anything in the forum)?

[cschneider]

We worked a bit more on this concept.
Our current idea is that agents live as Azure AI foundry agents. So they are deployed in the cloud and can be talked to via Azure agent sdk or ideally also via MCP.

The functions for the agents live as Azure functions and are accessed from the agent via MCP. The advantage here is that the function code as well as secrets for accessing your backends are encapsulated inside the function. So the caller does not need to know / care about these. The Azure agent can talk to the functions via MCP.

Autogen could then create small wrapper autogen agents that talk to the Azure agents.
So the agent coordination would live in autogen but the agents live remote in Azure. This way people could quickly define teams of already existing agents.

Ideally then the autogen teams could also live in the cloud and be addressed via MCP. So tools like chat bots could use existing teams without defining them.