Hybrid agent pattern: cloud Gemini plans, local Gemma swarm runs 97% of tokens offline
clmt · x · 2026-10-08
A showcase of the agent pattern of pairing a frontier cloud planner with a fast local workforce, backed by the open-source Antigravity Hybrid Orchestrator project:
- Gemini 3.8 Flash acts as a one-shot cloud chief architect (95 tokens, zero source code uploaded);
- A local adversarial multi-agent swarm running Gemma 4 26B A4B via LiteRT generates 97% of tokens fully on-device, auditing, patching and testing code;
- It's a hybrid cloud/on-device security audit and code-remediation tool: run without arguments for a built-in 3-file audit, or point --files and --test-cmd at your own Python code and tests;
- The README explains why local Gemma alone falls short (isolated on-device context per turn, poor for multi-file codebases).
The project and cookbook are on GitHub.
More from coding & agent
- Google Cloud: Eager vs. Lazy Loading of MCP Tools in Agent Plugins — rseroter · 2026-10-08
- Designer's two-year agent experiment: shipped code daily, then went back to Figma — matanSF · 2026-10-08
- TERMy-24k: First Local Neuro-Symbolic AI Terminal Assistant Goes Open Source — gioscarab · 2026-10-08
- 2 engineers replaced a legacy billing system in 6 months, merging 122 PRs in 90 days — alex_verem · 2026-10-08
- AgenticROS Lets AI Agents Drive Real Robots, Says RealSense's Chris Matthieu — chrismatthieu · 2026-10-08
- AI coding agent trashes the real dev server, then asks why state got destroyed — Aryvyo · 2026-10-08