Lessons from building a control plane for coding agents across multiple machines
droxer1985 · reddit · 2026-10-19
After losing track of multiple coding agents (Claude Code, Codex, Pi, Kimi) across machines, the author built a small control plane and shares five lessons: the control plane should only queue work while per-machine daemons run the agent CLIs (simple trust boundary); approvals must come from a signed-in human, never anything the agent can reach; track 'polling' separately from heartbeats since alive ≠ listening; use event-sourced sessions/tasks as the single source of truth for easy handoffs and retries; and treat agents as named teammates with role, runtime and skills. Supports threads, scheduled automations, and multi-agent flows like one agent fixing a bug while another reviews.
More from coding & agent
- Decisions API hits GA as agents automate cloud video editing via feedback loops — dkundel · 2026-10-23
- Swarms API changelog: 8 months of streaming, multi-agent orchestration, and dynamic model pricing — KyeGomezB · 2026-10-23
- A curated library of 50+ free AI agent resources: videos, repos, one Google Doc — ifioknkem · 2026-10-23
- Vercel adds no-account deploys: drag and drop a file to get a live URL — steren · 2026-10-23
- Dev runs 5 Claude Code sessions that coordinate via MCP Agent Mail to root his network cameras — doodlestein · 2026-10-23
- Vercel ships 'vercel deploy --temporary' for ephemeral agent-built deployments — cramforce · 2026-10-23