# The Simulation Layer Is Becoming Infrastructure

# The Simulation Layer Is Becoming Infrastructure

Shopify just open-sourced something more interesting than another model weights dump. Tangle, Tangent, and SimGym aren't tools. They're a statement about where production AI is actually heading.

The TL;DR from their recent Latent Space interview: Shopify now runs ~100% AI-assisted workflows internally. Their CTO Mikhail Parakhin (ex-Bing) described a phase transition in December 2025 when model quality crossed a threshold that made agentic tooling actually usable at scale. What's notable isn't the adoption curve — it's what they built to survive that curve.

**Tangle** is a content-addressed ML workflow engine. Think Airflow, but where every step is hashed and cached. If two researchers run the same preprocessing pipeline, the second one completes instantly because Tangle recognizes the content signature. This matters because ML experimentation has historically been a disaster of hidden state — notebooks that don't rerun, dependencies that drift, results that evaporate when the original author leaves. Tangle treats reproducibility as a systems problem, not a cultural one.

**Tangent** layers auto-research on top. It runs experiments, measures outcomes, and iterates. Parakhin mentioned 400+ experiments running automatically, with one success — but that one success was an optimization humans had missed in years of manual tuning. The shift here is subtle: instead of humans specifying hyperparameters, Tangent explores the solution space directly. AutoML finally works because LLMs can reason about *why* something failed, not just *that* it failed.

**SimGym** is the most interesting. It's a browser farm running simulated customers — multimodal agents that browse merchant stores, add items to carts, and behave according to distributions learned from actual Shopify transaction history. The historical data is the moat. Without it, you're just prompting agents to pretend they're shopping. With it, you can run counterfactuals: what happens to conversion if I change this banner? The simulation correlates ~0.7 with real A/B tests, which is close enough to screen obvious mistakes before they ship.

What connects these three systems is a realization: **the bottleneck in agentic AI isn't generation — it's verification at scale.**

Shopify's internal stats are telling. They fund unlimited tokens for employees but mandate Opus-4.6 or better. Token volume exploded post-December, but the distribution skews heavily — some engineers consume orders of magnitude more than others. The real metric Parakhin tracks isn't tokens generated. It's tokens spent on *review* versus *generation*. His ideal ratio skews heavily toward critique loops: one agent writes, another critiques with a different model, the first revises. Latency goes up. Quality goes up more.

This is infrastructure thinking. Most companies deploying agents are still treating them like faster interns. Shopify is treating them like a new runtime that requires entirely new CI/CD metaphors. When code is generated at machine speed, your pull request workflow becomes the bottleneck. They built their own PR review system because off-the-shelf tools assume human-scale throughput.

The SimGym piece is particularly instructive for anyone building multi-agent systems. Customer simulation isn't new — the Smallville paper, various gaming environments — but production-grade simulation requires something most don't have: ground truth data spanning years of real behavior. Shopify can validate their agents against actual purchase histories. Without that anchor, simulation is just expensive hallucination.

There's a broader pattern here. We're moving from "AI-assisted" tools to "AI-native" infrastructure. Tangle isn't a better notebook. It's a bet that ML workflows should be immutable, cacheable, and collaborative by default. Tangent isn't hyperparameter tuning. It's a bet that exploration should be automated end-to-end. SimGym isn't synthetic data generation. It's a bet that some domains are complex enough to require simulation *before* production, not just evaluation after.

The tools will matter less than the assumptions baked into them. Shopify is showing what happens when you stop asking "how do we use AI?" and start asking "what infrastructure does AI require?"

For builders: the simulation layer is coming. If your domain has historical data worth learning from, start there. Everything else is just prompting.