By skyforbes Recently I've been madly using Gemini as the main “brains” in a local multi-agent system for software development, and I’d love feedback from people here who are pushing Gemini beyond simple chat/IDE use.
The project is called QonQrete, and I’ve just released v0.5.0 beta.
What it is
QonQrete is a local-first orchestration framework where several LLM-backed agents cooperate to work on a codebase:
- InstruQtor – understands the task and produces a structured action plan
- ConstruQtor – applies that plan (editing / creating files, running tools in a sandbox)
- InspeQtor – reviews the changes and points out issues or follow-ups
Each role can use a different model/provider, but in my current setup Gemini 2.5 Pro is doing most of the heavy lifting for:
- repo understanding and task planning
- step-by-step reasoning over existing code
- explaining tradeoffs and risks in the review phase
Key properties:
- Local-first / self-hosted: The orchestrator runs on your own machine or server. You configure your Gemini API key (and optionally others) directly—no extra cloud service in the middle.
- Execution isolation: Any code or shell commands suggested by the agents run only inside containers (Docker/microsandbox), not on the host. QonQrete just collects logs and diffs.
- Human-in-the-loop or autonomous: You can either approve each stage manually or let it run full cycles on its own for faster iteration.
- Multi-provider support: Gemini can be combined with OpenAI, Claude, DeepSeek, etc., per role. I’ve found some nice combos where Gemini plans and reviews, and a cheaper model does the bulk coding.
Why I’m posting here
I’m specifically interested in how to get the best out of Gemini in this kind of setup:
- prompt patterns that make Gemini a strong planner or architect
- good ways to have Gemini critique/score diffs from other models
- ideas for tools/functions that pair well with Gemini in an agent loop
If you’re using Gemini for more “orchestrated” workflows or want to tinker with something like this, I’d really appreciate feedback, criticism, or ideas.
Repo (AGPL, work in progress):
https://github.com/illdynamics/qonqrete