He Named His AI Coworker MARVIN. It Runs 90% of His Day.
Sterling Chin didn't automate his job. He automated the parts that weren't worth his time.
Sterling Chin set a 40-day deadline to turn his AI assistant from a liability into a peer.
Not a tool. Not a chatbot. A system that handles meeting transcripts, Jira tickets, calendar management, and blog drafts, in sequence, with context carrying through from task to task, while Sterling focuses on the work that actually requires him.
He called it MARVIN. Now, MARVIN has nearly 900 GitHub stars. And watching it run live in this episode is different from reading about it, because the impressive part isn’t any single capability. It’s the handoffs.
Most AI demos show you one thing done well. MARVIN shows you what happens between things. The moment a meeting ends, the tickets update themselves. The moment a draft appears, before Sterling has thought to ask for one. That connective tissue is what separates a useful tool from something that changes how you work.
Treat your AI like a coworker
The mental model Sterling used to get there is worth stealing. He treated MARVIN like a new hire, not software. Written rules. Consistent feedback. Specific onboarding. Patience with early mistakes. He wasn’t configuring a product. He was managing someone through a learning curve, just like you would with a newly joined teammate. That approach produced a system reliable enough to trust with 90% of his day.
If you’re stuck with disappointing AI results, you probably skipped crucial onboarding. Context matters.
A few things from this conversation I keep thinking about:
The non-technical adoption story. Sterling rolled MARVIN out to 12+ colleagues at Postman, including people with no engineering background. They weren’t slower to adopt. They were faster. No legacy workflows to protect, and the value was immediate: their jobs are mostly text-based coordination work, which is exactly what MARVIN handles best.
Sub-agent architecture. Without defined personality rules and explicit constraints, AI assistants default to being generic. MARVIN has a specific tone, style, and task scope built in. Those constraints arguably make it trustworthy. Flexibility without structure produces inconsistency, and inconsistency is why most people abandon these tools.
The DIY vs. big tech gap is wider than most people expect. Big tech assistants optimize for everyone. MARVIN is optimized for Sterling’s exact workflow. The difference between “good enough broadly” and “right for me specifically” is significant. Open source on top of a capable model closes it.
This episode is worth watching rather than just listening. Sterling demos MARVIN live, and the screen shares matter.
Also on Spotify, Apple Podcasts, or wherever you listen.
But while what Sterling built with MARVIN is impressive, it also touches something that makes a lot of people uneasy. Replacement.
Not Sterling replacing himself, exactly. Instead, the question underneath it: if an individual engineer can build an AI that handles 90% of his day in 40 days, what does that mean at scale?
I’ve spoken before about AI agents as ‘async junior digital employees’, and written about the rise of tireless digital workers. The individual version of this story, one person extending their capacity, building something that amplifies their work, feels exciting.
But what happens when it goes far enough that the agent can actually replace the role? Not extend it. Replace it.
We’ve all consumed that kind of dystopian science fiction.
Now, some CEOs are trying it for themselves. The results are complicated.
I’m writing about what that looks like next week. Warts and all.
See you then.
- Conor