Monaco published an account of moving to cloud-hosted agent-developer workspaces, and the valuable half is the failure. Its original Million Interns idea was a manager agent spinning off dozens of intern agents, assembled from the tools at hand: a session pane per agent, iTerm2 notifications when input was required, a fresh Git worktree, a tmux session. A couple of engineers ran it for a while and it was never adopted company-wide.
The stated reasons are concrete. Tmux is not a universally loved interface for developers to work in. Git worktrees are flimsy and need setup each time — pre-commits, virtualenvs, package installs. And there was no elegant way to notify the human in the loop when an agent needed an answer. None of those is a model problem; all three are workspace problems, which is why the answer became hosted workspaces rather than better prompts.
What made it urgent was adoption: Monaco saw a swell of Claude Code usage in late 2025 and early 2026 around the Opus 4.6 release, with developers running agent sessions in parallel across worktrees and learning to context-switch between them.
Worktree isolation surfacing a control-plane problem is exactly the finding in Orca solving agent workspace isolation and then surfacing the next problem, and it is the same trajectory as local coding agents becoming shared workspaces. The agent harness hub tracks it.
This is one company's engineering narrative with no cost, throughput or adoption figures attached to the hosted approach. The condition to watch is whether Monaco publishes what the hosted workspaces cost per developer, because the original attempt failed on ergonomics and the replacement will be judged on price.