Shopify's Head of Engineering on what happens after the AI writes the code: how people, review, and the org get rearranged
What Is Your Job Now, Farhan Thawar | Compile 26 · Farhan Thawar
24 min total·Actually worth watching closely: ~6 min·3 must-watch clips
- 0:00 – 2:31Watch
The era of hand-written code is over
He opens with the conclusion: all the software we've come to use was written character by character, and that era is finished. Then the shift in where time goes across the development lifecycle — the coding step shrinks, and planning, analysis and validating AI output become the bulk.
The question isn't whether AI can write code; it's what engineers, and the whole R&D org, are actually for once the coding step collapses.
This stretch has a before-and-after chart of where time goes in the lifecycle, and he points at it while he talks — the two columns make the shift obvious at a glance, and audio alone loses that structure.▶ Jump to 0:00Speaker · Farhan Thawar - 2:31 – 7:01Listen
The new way to write code, and the scarce skill
Two new shapes of coding: parallel — split the task into subtasks across agents, with the human consolidating and validating; and sequential — pair deeply with one high-reasoning model, then bring another model in to interrogate it. Then value shifting to taste and judgment: when to build, why to build, how to trade off architecture.
The question that tests a proposal is "how did you choose this solution among the 10,000 possible right solutions?" — being able to answer it is what judgment means.
Pure argument here: he's mostly describing how he works and what he asks in reviews, with the slide parked on a title, so listening while walking or commuting costs you nothing.▶ Jump to 2:31Speaker · Farhan Thawar - 7:01 – 12:11Skim
From first Copilot users to a baseline expectation
Shopify's timeline: they were the first users of GitHub Copilot in 2021, a year before ChatGPT — but the real turning point was the CEO's internal memo that AI would be a baseline expectation, after which tool usage rose across the board and a lot of peer CTOs and CEOs copied it.
An org turns on two things at once: an unambiguous statement from the top that leaves no way back, and tools underneath that genuinely work. Neither does it alone.
The middle is company-history narration you can skip ahead through, but stop at the memo to look at the usage curve — the rise after that date is the only hard evidence in this stretch.▶ Jump to 7:01Speaker · Farhan Thawar - 12:11 – 14:03Listen
The first ones using it weren't engineers
1,500 Cursor licenses were gone in a week, taken by finance, HR, sales and marketing — they treated a coding tool as their way into LLMs and agents, and built real things: a feedback query tool for support, business reviews for customer success, forecasting tools in finance.
Demand for AI tools outside engineering is badly underestimated; handing coding tools to those teams turns the company into a bottom-up culture engine.
All spoken examples, department by department — no visuals, and none needed. Nothing is lost by listening only.▶ Jump to 12:11Speaker · Farhan Thawar - 14:03 – 17:40Listen
Token culture, and big models only
The goal moves from reaching for AI reflexively to using AI for leverage, and burning tokens becomes something to brag about. No spend caps, just circuit breakers: when something looks runaway, a confirmation in Slack lets it through. And engineering is explicitly not allowed to use smaller models.
The math behind banning smaller models: human time is worth more than clanker time, and the hours spent hunting one small bug a smaller model introduced cost far more than the token difference.
This is a management argument — the slide is a few policy lines, and the value is entirely in the reasoning chain and cost comparison he says out loud, so listening and taking notes is enough.▶ Jump to 14:03Speaker · Farhan Thawar - 17:40 – 21:05Listen
Roadmaps in weeks, learning as the asset
If the AI is writing the code, a six-month roadmap deserves the question "why not this week, why not today?" Then the most counterintuitive point of the talk: code isn't the collateral, learning is — the CEO once had a 50-engineer team delete 18 months of work and rebuild on a simpler architecture, and it took three months. It closes on why pair programming is still worth keeping.
Two questions to put to your team every week: what did you learn, and what secret do you believe that others don't?
The deleted-18-months story and the case for pair programming are narrative and argument — he delivers them standing there with nothing extra on screen, so listening is the efficient way through.▶ Jump to 17:40Speaker · Farhan Thawar - 21:05 – 22:40Watch
LGTM and the council of LLMs
Once AI floods you with code, the bottleneck moves to review. He puts "the four scariest letters in the human language" on screen — LGTM — and argues it proves humans were not looking at the code to begin with. Shopify's replacement: several models reviewing different aspects and jointly judging whether it ships.
Model review isn't perfect, but it's better than humans were, and it compresses review from 24-48 hours to about an hour.
The four letters fill the screen and the room cracks up, then comes the diagram of how the review is divided among the judges — the joke and the mechanism are both on screen, so audio alone misses the laugh and can't reconstruct the system. Densest stretch in the talk.▶ Jump to 21:05Speaker · Farhan Thawar - 22:40 – 24:00Listen
Responsibility stays with people, and how to hire
The closing line in the sand: Shopify does not let AI take responsibility — you can have it write as much code as you want, but your name goes on the PR. Plus the intern class scaled from 75 to 1,000, bringing in a generation that has lived inside ChatGPT to question everything about how the company works.
"Responsibility can be shared, but it cannot be given away" — the one bolt you can't loosen once AI is everywhere.
It ends with a few plainly stated conclusions delivered straight to the audience, away from the slides; hearing the line itself is the whole thing.▶ Jump to 22:40Speaker · Farhan Thawar