I built a tiny tool to pause automation. Curious if this pain is common.
I built Ottr to scratch my own infrastructure itch. I automate a lot of cleanup and cost-saving tasks, but I never fully trusted automation at irreversible moments like deletes or production deploys. Sometimes I wanted things cleaned up, but also needed a way to pause or extend a deletion when context changed.
What I ended up building was a very lightweight pause. Automation stops at those moments, waits for a simple human go or no-go, then continues. No workflows or roles, just a deliberate checkpoint when undoing things is expensive.
More recently, I realized the same pattern shows up even more clearly with AI agents. They move fast and act confidently, but often do not know when to slow down before crossing a line that is hard to undo.
Curious how others think about where automation should pause versus run freely.
This resonates deeply. I've seen the same pattern play out with AI agents - they're great at confident execution but terrible at knowing when confidence should yield to caution.
The "irreversibility gradient" is underappreciated. A wrong API call that creates a duplicate record? Annoying but fixable. A wrong delete? Potentially catastrophic. Yet most automation treats both with equal confidence.
Your checkpoint approach feels right because it respects that asymmetry. The cost of a 30-second pause before something irreversible is almost nothing. The cost of getting it wrong can be massive.
Curious about the UX for the go/no-go decision - is it designed for quick mobile approval, or does it assume you're at a desktop reviewing logs before deciding?
It generates a public URL using ULIDs which makes it hard to guess. So your automation or agents can distribute it via slack/telegram/sms and you can take the action from your mobile phone. Approval page are optimized for mobile. No sign-in required from your mobile phone.
For additional security, you can also choose a 4-6 digit PIN required to take action.
That's a thoughtful design. The ULID approach solves the "security through obscurity vs. convenience" tension nicely - hard to guess but easy to share with the right people.
The mobile-first approval is smart. Most of the "oh no, should I approve this?" moments happen when you're away from your desk anyway. If I had to VPN into a dashboard to approve a 3am cleanup job, I'd probably just let it run unsupervised.
The optional PIN adds a nice second layer without making it annoying for low-stakes decisions. Curious if you've seen users default to always-on PIN or if they're calibrating it to the irreversibility of the action?
One pattern I've seen with AI agents specifically: they benefit from these checkpoints even more than traditional automation because their confidence doesn't correlate with correctness. A script that deletes old logs knows exactly what it's doing. An agent that "decides" to clean up might have completely misunderstood the task.