The work should outlast the laptop window
Development machines were assigned to individual people. Nirvaan and his director reorganized eight of them into shared resources, with project environments and agent sessions that people could reconnect to. The work could continue on its host after the operator disconnected.
Separate access, execution and memory
They chose Omarchy, from DHH/37signals, to reduce operating-system overhead. Nirvaan helped set up SSH access through separate Linux users and sandboxes. Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) kept agent sessions available across machines, while Batman, a lead agent running on Hermes (the agent runtime, not a language model), dispatched work. Hindsight retained longer-term working knowledge.
The graphical access idea came from his game-streaming habit. Sunshine and Moonlight let people share a machine's logged-in desktop and input. For machines that were asleep, a reachable NAS on Tailscale sent Wake-on-LAN packets over the local network. SSH users and sandboxes handled separate working environments; the streaming connection handled the shared desktop.
Keep sessions available after disconnects
Herdr (a third-party terminal workspace manager for AI coding agents; ARDA is Nirvaan's open-source plugin for it) keeps agent processes running on an awake host after the operator closes the interface. A person can reconnect to the work instead of keeping one laptop window open for the duration. Wake-on-LAN brings sleeping machines back into reach. Application workflows handle recovery separately, including PostgreSQL-backed continuation in GTM and Temporal recovery in FosterX. FosterX is CloudRedux's organisational AI platform, built by a team; Nirvaan is one of its primary engineers.
At CloudRedux, his production operations cover AWS and Linux deployments, Docker Compose, Nginx and CI/CD. His CloudRedux work includes swap fallback for memory spikes, AWS CloudWatch monitoring, logs and metrics, and uptime alerts to Discord and Slack.
Keep authority with people
Separate Linux users and sandboxes give people and agents defined working environments. Privileged access remains a deliberate operational choice. People retain architecture and consequential decisions, while agents execute delegated work and return results. Keeping access, dispatch, memory and communication separate makes each responsibility easier to reason about.
Shared development at scale
Designers and engineers spent less time re-establishing project context for their agents and more time on the work itself. Across shared development activity from June to September 2026, the setup recorded 2,420 agent sessions, 30+ concurrent agent sessions at peak (agent terminals, not people), 55+ subagents on one request, eight language models and 213 unattended agent runs. The eight-machine development environment gave people and agents a shared place to continue work across sessions.
Proposed next iteration
The next iteration would add a repeatable host-failure exercise and clearer scheduling visibility: show which work is waiting, which host owns it and what happens when that host becomes unavailable. Measuring contention under a defined workload would help decide when to reuse capacity and when another machine is justified.