Rube Goldberg AI Workflows
Musings + AI Generated Art, by Rahul Parundekar. Published on 12th May, 2025.
Image generated using GPT 4o on May 12th, 2025.
You folks know what a Rube Goldberg machine is, right?
As I implement at-scale complex multi-step workflows for my clients, I’m reminded of how the non-deterministic nature of LLMs compounds with multiple downstream steps. Combine that with standard engineering realities—rate limits, flaky connections, retries, timeouts—and it quickly becomes more than a toy problem. It becomes real engineering.
Take one of my current workflows: it takes 15 minutes to complete. That’s not idle time—it’s churning through PDF parsing, database reconciliation, structured-to-unstructured transformations, all of which carry their own failure probabilities.
At one point, I parallelized tasks using ThreadPoolExecutor, but rolled it back after too many hard-to-debug errors. Turns out, a pesky 503 from Bedrock was derailing things mid-run.
At this scale, you’re not just prototyping—you’re building distributed, fault-tolerant systems. AI workflows need the same care we give to production infra.
So, yeah… the comic above? It’s a little funny, but also a little too real.
All this… from one reply. Welcome to the future of autonomous growth.
🛠️ Curious how to make your AI agents production-ready? Drop me a note.