How to manage agents as different services that can interact with each other #2158
-
|
Use Case I'm trying to determine what is the best architecture for my use case. Basically I will have an orchestrator agent that will have a lot of subagents (maybe somewhere close to 50). There will also be a lot of MCP servers that will be available to those subagents. The orchestrator agent will need to be able to use any of those subagents to complete different tasks. The difficult part is that the orchestrator agent should be able to dynamically load what subagents are available to them, and each subagent should be able to dynamically load what MCP servers are available to them. Proposed Architecture I could deploy each adk agent and each MCP server as its own container/service in Cloud Run. There would be a main orchestrator service (we can figure out if there needs to be another layer of subagents under this) that can dynamically load what agents are available from Firestore. Firestore would contain all of the metadata for the different agents/deployed services and MCP servers that are available, so you would just need to make a change here if you were adding/removing agents. If you need to edit a single agent or MCP server, you only need to redeploy for that agent/server. And if one agent isn't working/available, it doesn't disrupt the whole task. Agents can dynamically load what MCP servers are available to them (once again using Firestore). As for subagents that need to pass a task over to another subagent - I guess the remote subagents available to a subagent could also be made dynamic. But to me this doesn't seem like real A2A? I thought A2A had to be agents talking to each other in a single ADK app, not remotely accessing different Cloud Run services. Maybe this is all complete overkill but I've never created a multi-agent architecture of this scale. I've only made monolithic adk projects where everything (orchestrator + subagents) were managed in one Cloud Run container. Does this solution seem scalable? I'm also wondering if Vertex AI engine can do something similar to what I'm proposing with Cloud Run, I'm not sure I quite understand how the engine is used/how code changes are made. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments
-
|
Any update on lessons learned here @benoudrih ? |
Beta Was this translation helpful? Give feedback.
-
|
@benoudrih @paulrinaldi skip to 8:40 in the video to see him show a code example where he's using adk. |
Beta Was this translation helpful? Give feedback.
@benoudrih @paulrinaldi
check this video
https://www.youtube.com/watch?v=_Oczktevufo
skip to 8:40 in the video to see him show a code example where he's using adk.
expand the Architecture from the video to your subagents(probably using A2A) as well as your tools