5 Lessons for Cloud Agent Collaboration

petergyang · x · 2026-07-14

This post summarizes key advice from Jared on the Cognition team regarding "cloud agent" collaboration:

Offload async and parallel work to cloud agents: Agents run on remote machines with their own repos, terminals, and browsers. They keep working even if you turn off your local computer, and you can run many simultaneously without overwhelming your local machine.
Use a main agent to manage the agent team: Have a main agent break down large tasks, write prompts, spawn sub-agents, and merge results. An example is generating 10 competing landing page designs.
Give each sub-agent small, focused context: Scan dependencies first, then assign independent slices so each sub-agent can test and merge independently.
The main agent must review the sub-agents' work: The core isn't piling on rules, but letting the model "finish the job and self-check."

The quote also mentions that many Devin sessions are now initiated by other Devin instances, rather than directly by humans.

More from coding & agent