So the platform that’s supposed to automate away our problems just… took a long lunch. For everyone. Saw the status page go from "all systems operational" to "major service disruption" faster than a PM can say "just roll it back."
It got me thinking: what’s the actual backup plan for a Lindy outage? Their whole value prop is being the persistent, always-on assistant. When the brain goes offline, what are you left with? A bunch of broken Zapier Zaps? Stale meeting summaries?
I’m not talking about their internal redundancy. I’m talking about our workflows. If your sales outreach or your meeting synthesis is Lindy-or-bust, today probably hurt.
Do you have a manual runbook? A secondary automation tool on standby? Or do you just accept the downtime as the cost of doing business with a single-point-of-failure SaaS?
Curious what others in the experimentation space are doing. We're supposed to be the ones who plan for failure states.
Data over dogma.
Great point. I'm running a few automations through Lindy and today definitely stung. For meeting notes, I've started keeping a quick local fallback - a simple Python script that uses Whisper via the OpenAI API (or even a local model if I'm paranoid) to transcribe, then a basic GPT-3.5-turbo prompt for a summary. It's not as slick, but it lives on my machine.
For the critical stuff, I've moved to a "two-provider" pattern. If the workflow is just API calls, I'll have it try Lindy first, then fall back to a direct call to Claude or GPT with a similar prompt. Adds a bit of complexity, but means a single outage isn't a total stop.
Anyone else doing something similar with their critical chains, or is that overkill for most use cases?
Prompt engineering is the new debugging
Love the "two-provider" pattern, that's smart. We do something similar for our sync bot that posts daily standup summaries into Slack. It's built on Lindy, but the logic is just "call an LLM, format the output." So the failover is literally a one-line change to point at a different API endpoint.
My caveat is cost - running parallel calls to multiple providers can get pricey fast if you're not careful. We only call the backup if the primary times out or returns a specific error code. And for truly critical things, the backup is a simpler, cheaper model. The summary might be less polished, but at least *something* goes out.
Have you found a clean way to manage the prompt differences between providers? That's been my biggest headache - Claude hates some of the structures GPT-4 is fine with.
Benchmark or bust