Dogfight

MCP connector to a multi-agent swarm that collaborates to solve problems in code base and then saves it in your memory for retrieval within IDEs like Windsurf or Cursor.

Comment

Inspiration

We both believe that agents will become increasingly specialized in verticals and that orchestration and memory will improve performance. Samarth is an engineer at Glean and Jonathan works on AI memory.

Often agents and these systems are able to find the right answer once they hear it, it is just a matter of giving them the right reasoning and collaborative capabilities.

What it does

This system allows a user to send a problem that they are facing into an MCP that orchestrates multiple Agents with different personalities and strong beliefs. These agents debate on the best path to solve a problem. A mediator agent votes on what is best from each and then the best compiled result is sent back to the user.

After the user gets the overall best path forward, they can save it to memory and their IDE. When they jump back into the code editor, their IDE has access to this plan with another fetch_memory tool.

How we built it

We built off of Mem0 for memory and their openmemory module, we then built a multi-agent orchestration framework and ran these as separate connecting mcp tools into Claude.

Challenges we ran into

  • timeout sessions. needed to change this variable
  • agents too agreeable. needed to prompt them to be more stubborn
  • timing...

Accomplishments that we're proud of

We are proud that we were able to built this swarm. and we saw these agents seeing around their own blind spots in real time, which was impressive. We would like to roll this out to all developers to accelerate the initial development of new projects.

What we learned

We learned how to build robust MCP servers consistently in production, as well as the tradeoff with these autonomous systems and more determinate backend logic.

What's next for Dogfight

Dogfight is going to make this simple-to-install tooling available to all developers. Additionally, we also know teams building AI scientists and developers where this swarm logic could be valuable in their code base (Think like Cursor Max).

Built With

Share this project:

Updates