Your AI Doesn't Need More Uptime. It Needs a Second Door.
TL;DR: My production service survived four separate waves of losing SSH access without going down once. What failed, four times, was the only door I had into the machine. Uptime protects the service. A second door protects your ability to run it — and for a one-person operation, the second is the one that matters. Four times, my management access to my own production machine died. Four times, the…
In the world of AI systems, uptime is often emphasized as the primary metric of success. However, Jiahui Miao, the founder of vertciti, argues that a second door is equally, if not more, crucial for the proper functioning of a one-person operation. She shares her experience of losing SSH access four times without any impact on the service itself, while the proxy responsible for the SSH connection kept cutting port 22.
This example highlights the importance of having an alternative access method, independent of the primary access channel.
Miao emphasizes that uptime protects the service, but a second door safeguards the ability to manage and operate it. A single point of failure in the access path can be disastrous, as users would remain unaware of any issues if the primary and secondary access methods fail together. She stresses the need to differentiate between a dead service and a dead path, using an independent health signal to confirm the status of the service.
To build a reliable second door, Miao outlines the essential components. First, there should be an endpoint for health checks, separate from the primary access method. Secondly, a management channel that operates independently of the primary access channel is necessary. In her case, she used the provider's run-command service, which dispatches commands to the agent running on the instance without relying on SSH or facing the same fate as the primary access.
However, Miao warns that building a second door is not a trivial task. It requires careful testing and end-to-end verification of all three components - the instance agent's run-command plugin, the IAM policy granting authorization, and the execution user on the machine with proper permissions. This process took several hours for Miao, as a single missing piece can render the entire setup ineffective.
Miao's key takeaway is that uptime is a property of the service, while access is a property of the operator. For agent-run systems like hers, the critical factor is not the service's availability but the operator's ability to independently observe, diagnose, and act upon the system when the primary path fails. Redundant replicas and status pages cannot make up for a lack of a second access method during critical incidents.
In conclusion, Miao argues that a second door is not just an added bonus but a necessity for ensuring the smooth operation of agent-run AI systems. By building an independent health signal and a robust management channel, operators can maintain control over their systems, even when faced with unexpected access issues.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
