How I Run AI Agents in the Cloud Without Leaving Claude Code (with Aether)
DevOps & AI Toolkit · 2026-10-19
💡 Quick Take
1. The host offloads agent workloads to Aether, a cloud service that runs coding tasks in disposable microVMs (lightweight virtual machines).
2. Aether now lets developers trigger cloud runs directly through terminal agents like Claude Code, avoiding web dashboards entirely.
3. According to the host, Aether's tool integration currently lacks machine sizing options, locking every cloud task to one default compute tier.
4. Status alerts failed to reach the host's main session overnight, forcing him to check GitHub emails to confirm finished work.
5. The host advises keeping daily steady workloads on owned hardware and renting cloud microVMs only for occasional burst capacity.
📊 Detailed Explanation
The host uses a dedicated home box running his custom tool, Agent Deck, to manage coding agents across separate git worktrees (isolated project working folders). His setup easily handles steady daily tasks. The problem starts when five or six agents compile code or run test suites at the same time. The machine runs out of processor cores, memory, and disk speed. Upgrading hardware for spikes that happen twice a week makes little financial sense. The host wanted extra capacity without leaving Claude Code or Open Code for a web dashboard. Aether solves this by spinning up a separate microVM for each assigned task and tearing it down when the run finishes.
During an overnight test, the host sent ten tasks covering twelve issues to Aether while running nine local agent roles on his own hardware. The remote tasks completed cleanly, producing verified pull requests. However, several rough edges surfaced. The background check in his terminal failed to register the updates, so his dispatching agent never knew the jobs finished. The dispatch tool also hid virtual machine size controls from the agent. Claude Code understood the repository build requirements, but it could not adjust the machine power for heavy tasks. According to the host, developers still act as manual schedulers because no single tool coordinates local agents and cloud instances together.
🎯 Tech Expert Opinion
The host's hybrid setup provides a practical blueprint for handling AI workloads. Running local coding agents that build real code exhausts standard developer machines quickly. Buying top-tier hardware that sits idle most of the week is an outdated strategy. According to the host, offloading peak compiles to transient cloud microVMs matches how modern infrastructure handles irregular traffic. Developers keep their familiar terminal setups while borrowing heavy compute only when queues back up.
The biggest bottleneck in this architecture is the broken feedback loop. The host's testing exposed that remote agents cannot reliably report state changes back to the primary orchestration loop. In agentic workflows, an orchestration agent needs to know the exact moment a coder agent finishes so it can trigger reviews, tests, and pull request merges. When notifications fail silently, autonomous pipelines grind to a halt. Remote agent platforms must prioritize robust webhooks and bi-directional event streaming over fancy web interfaces.
Security and process isolation present another key trade-off. Running untrusted or experimental code in disposable microVMs protects developer laptops from greedy test loops and disk pollution. At the same time, this workflow requires sharing code repositories and broad authentication tokens with external cloud platforms. Teams handling sensitive corporate codebases must weigh the benefits of instant compute scaling against the security cost of exposing repository credentials to third-party runners.
According to the host, bursting overflow into cloud microVMs makes sense for solo developers and small teams hitting hardware ceilings during heavy build bursts. Developers who require instant, single-keystroke visibility into every running process or operate under strict on-premises data policies should wait until fleet management tools offer better bi-directional sync and local oversight.
Kanal: DevOps & AI Toolkit