Fresh context per agent, shared branchable Git repo to compound discoveries
anirudhg9119 · x · 2026-10-17
The author shares a multi-agent experimentation approach with Ethan Mollick: each agent gets a fresh context, while past experiments (including failures) are preserved in a shared, branchable Git repo.
The goal is fresh perspectives without context pollution, plus compounding discoveries across agents via the repo. Context: Mollick had pointed to a paper where a team of Gemini 4 agents beat Minecraft, showing spontaneous coordination among agents.
More from coding & agent
- hitslop: open-source single-file mini apps on macOS that your AI agent builds and edits — Jonovono · 2026-10-17
- Sentry CEO: sandbox workloads run on Vercel, internal services on Northflank — zeeg · 2026-10-17
- Seroter's reading list: context inertia, agent evals past the vibe check, 3x code shipping — rseroter · 2026-10-17
- Muxin: open-source desktop app to run AI coding agents across all your machines — sull · 2026-10-17
- HeyPCB launches MCP server letting AI agents design PCBs and export KiCad projects — R4nsack · 2026-10-17
- Garry Tan shows agents coding his new project and handling DevOps in one go — garrytan · 2026-10-17