Fetching from the wire…
01
02
03
04
05
06
07
08
Top 5 · 2026-04-23 · source-backed
Mikhail Parakhin doesn't do half-measures. In a Latent Space deep-dive interview, Shopify's CTO (ex-Microsoft, ex-Bing) revealed that 100% of Shopify's workforce now uses AI daily, and the company actively discourages anyone from using a model less capable than Opus 4.6. Not recommends. Discourages.
The budget policy is wild: unlimited tokens for everyone. No approval process. No department caps. Parakhin's logic is that the marginal cost of tokens is so low relative to employee time that restricting them is penny-wise, pound-foolish. I've been running my own pipeline with token budgets and I still flinch at the monthly bill, so hearing a $200B+ company say "just spend it" caught me off guard.
But the really interesting part isn't the spending. It's the quality metric. Parakhin tracks the ratio of generation tokens to automated review tokens. For every piece of AI-generated output, a high-end model reviews it. The critique loop, not the generation step, is where quality lives. He's basically saying the cost of generating is noise. The cost of validating is the investment.
This tracks with what I'm seeing in my own work. I spend more time reviewing and steering AI output than I do prompting for it. The bottleneck moved months ago from "can AI write this code" to "can I tell whether this code is right." Shopify is formalizing that intuition into a budget line.
Then there's SimGym. Shopify built an internal environment for training AI agents on simulated merchant scenarios before deploying them to real stores. Think of it as a staging environment, but for agent behavior rather than code. Agents learn merchant workflows, edge cases, and failure modes in simulation. When they graduate to production, they've already handled the weird stuff.
The Tangle and Tangent products he mentioned deserve their own write-up, but the SimGym pattern is the one that generalizes. If you're deploying agents that interact with users or customers, build a simulation layer first. Let the agent make mistakes where they're cheap. This is how Shopify gets to 100% adoption without 100% chaos.
What builders should take from this: stop rationing tokens. The generation-to-review ratio is a better quality metric than any benchmark. And if you're serious about agents in production, build your own SimGym. The simulation layer is where confidence comes from.
Each link below shares sources, entities, or timing with this story.
Mikhail Parakhin works at Shopify / Shared entities / Same source / Shared topic / What happened next
Linked by a graph relationship (Mikhail Parakhin works at Shopify); both cover Latent Space, Opus, Parakhin, Shopify; cite the same source (Latent Space deep-dive interview).
Linked by a graph relationship (Mikhail Parakhin works at Shopify); both cover Latent Space, Opus, Parakhin, Shopify; cite the same source (Latent Space deep-dive interview).
Mikhail Parakhin works at Shopify / Shared entities / Shared topic / What happened next / Tension
Linked by a graph relationship (Mikhail Parakhin works at Shopify); both cover Opus, Shopify, Then, When; overlapping topics (agent, token).
Claude Code uses Opus / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Code uses Opus); both cover Opus, Then, When; overlapping topics (agent, cost, quality).
Mikhail Parakhin works at Shopify / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (Mikhail Parakhin works at Shopify); both cover Latent Space, Shopify; reported by the same outlet (latent.space).
Claude Code uses Opus / Shared entities / Shared topic / What happened next
Linked by a graph relationship (Claude Code uses Opus); both cover Microsoft, Opus, Then; overlapping topics (agent, code).
Shopify uses Claude Opus / Shared entities / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Shopify uses Claude Opus); both cover Latent Space, Opus; reported by the same outlet (latent.space).
Claude Code uses Opus / Shared entities / Shared topic / What happened next
Linked by a graph relationship (Claude Code uses Opus); both cover Opus, When; overlapping topics (agent, code, cost, token).