Hi everyone,
I recently finished the alpha-stage development and my first full week of real-world testing for a project called Backplaned. I wanted to share it here because the Gemma models were central to making the architecture actually work in practice.
What is Backplaned?
It’s a lightweight, self-hosted multi-agent orchestration platform. Instead of a monolithic app, it uses a central router acting as an ESB-style message broker. It handles task routing, access control, and file transfers between pluggable agents over a unified HTTP protocol.
It ships with a suite of tools including a sandboxed code execution workspace, long-term vector memory, document-to-markdown conversion, and MCP bridges.
How Gemma Powers the System
For my baseline foundation model during testing, I deployed Gemma 4. In a multi-agent setup like this, the core orchestrator model carries a heavy cognitive load. It has to act as the central “brain,” evaluating user intents, deciding which external agent or tool to route tasks to, and synthesizing the final output. Actually, all examples in showcase are powered by Gemma 4 26b a4b!
I was incredibly impressed by how well Gemma 4 handled these complex agentic tasks:
-
Tool Calling Reliability: It consistently adhered to the strict tool schemas required to trigger external agents (like kicking off a DuckDuckGo search or executing a python script) without hallucinating parameters.
-
Context Maintenance: It easily maintained the context window through multi-step research and execution loops, synthesizing the results from various agents back to the user seamlessly.
-
Local Inference Efficiency: Serving as the centralized LLM gateway for the whole network of agents, it was highly responsive and efficient as the baseline model.
If you are working on multi-agent workflows, tool-calling loops, or local AI deployments, I’d love to hear your thoughts on the routing architecture or your own experiences using Gemma for complex orchestration!
Repo link: https://github.com/SJK-py/backplaned