Domenic's agentic coding setup: disposable Linux VMs and blast radius over model choice
bibryam · x · 2026-10-04
Domenic Denicola details his July 2026 agentic coding workflow: coding agents run with broad autonomy inside disposable Linux VMs, accessed via Tailscale, isolated with git worktrees, and saved to GitHub — letting him fix production bugs from his phone on a train.
Key insight: the design choice that matters is the blast radius, not the model.
Requirements and design points:
- Agents must run on Linux: they're trained heavily on Bash/Unix tools and trip over quoting and UTF-8 on Windows
- Seamless desktop/laptop handoff: preserving agent sessions, working tree state, and gitignored files (.env, build artifacts)
- Work continues after closing the laptop
- Global settings (AGENTS.md, skills, harness configs) version-controlled and available everywhere
- Review agent work live in VS Code, not just harness UI or GitHub PRs
- Agents can stand up dev servers for review
The result is a complete autonomy-first agent infrastructure built around safety boundaries.
More from coding & agent
- Antigravity Rolls Out Claude Opus 5.5 and Sonnet 5.5 to All Paid Users — firstadopter · 2026-10-04
- Elden Ring × Super Mario 64 mashups go viral, built with real game code via AI coding models — mark_k · 2026-10-04
- OpenAI quietly ships MCP Events, bringing event-driven triggers to agents beyond cron — nbaschez · 2026-10-04
- Codex remote beats Claude Code and Cursor, dev says after hands-on comparison — Angaisb_ · 2026-10-04
- Mediabunny's seeking logic wins praise for fixing Chrome's broken variable-frame-rate playback — Vjeux · 2026-10-04
- PrAIvy: A P2P Network to Share Local Ollama Instances Across Users — Matty_za33 · 2026-10-04