Shopify’s AI Phase Transition: 2026 Usage Explosion, Unlimited Opus-4.6 Token Budget, Tangle, Tangent, SimGym — with Mikhail Parakhin, Shopify CTO

Episode Summary

Executive Summary: Shopify CTO Michael Parkin described how AI is reshaping internal engineering, research, and commerce workflows: near-universal employee AI adoption, a shift toward critique-heavy coding agents, and major gains from new systems like Tangle/Tangent for experimentation and SimGym for customer simulation. He argued that larger models used more deliberately can reduce bugs and improve deployment quality, while Shopify’s moat is increasingly built on proprietary data and agentic infrastructure.

Main Topics: Enterprise AI adoption at Shopify (Priority: 5/5): Parkin showed internal charts indicating AI tools are now used by nearly everyone at Shopify, with the biggest growth coming after a model quality inflection in December. He contrasted CLI-based tools with IDE-based tools and emphasized unlimited token funding with model guardrails. Token usage, code generation, and quality control (Priority: 5/5): He argued that raw token consumption is less important than how tokens are used: the best pattern is fewer agents, stronger critique loops, and expensive models for review. He warned that more AI-generated code increases bug risk unless PR review becomes much stricter. CI/CD and repository workflows under AI pressure (Priority: 4/5): Parkin said traditional PR, merge, and CI/CD systems are becoming bottlenecks as code volume grows. Shopify is using stack-based workflows and Graphite, but he believes the industry may need a new paradigm for agentic software development. Tangle and Tangent for reproducible experimentation (Priority: 5/5): He presented Tangle as a platform for collaborative, reproducible data/ML workflows with content-based caching and easy promotion from experimentation to production. Tangent sits on top as an auto-research loop that can run experiments, iterate, and optimize objectives autonomously. SimGym for customer and merchant simulation (Priority: 5/5): SimGym uses Shopify’s historical behavioral data to create simulated customer/merchant agents that can predict outcomes of site changes, discounts, or interventions. Parkin framed it as a major moat because it improves with more data and can model counterfactual journeys over time. Liquid neural networks and low-latency inference (Priority: 4/5): Parkin said Shopify is actively using Liquid AI models for some workloads because they are efficient for low-latency, long-context tasks and can outperform other small models in specific production settings. He described them as a strong distillation target rather than a frontier-model replacement. Historical context and future hiring needs (Priority: 3/5): He reflected on Sydney as an example of intentional personality shaping in chatbots and discussed hiring needs in ML/data science and distributed databases, hinting that AI will change database architecture too.

Key Arguments: AI adoption at Shopify is now close to universal; the green line nearing 100% indicates employees are deeply integrating AI into daily work. CLI-based agent tools are growing faster than IDE-centric tools because they reduce friction and fit more agentic workflows. Token count alone is a poor management metric; better measures are critique-loop quality, model choice, and the ratio of generation to review spending. Using more tokens can improve outcomes if those tokens are spent in high-quality critique and PR review, even if latency rises. AI-generated code increases the total amount of code, so bug rates can rise even when model-written code is individually better than human-written code. Current PR and CI/CD systems were designed for humans and are straining under machine-speed development; a new workflow layer may be needed. Tangle solves a real pain point: reproducible experimentation and easy promotion from notebooks/experiments to production-ready pipelines. Content-based caching and shared execution can prevent duplicated work across teams and surface when the same computation is being rerun elsewhere. Tangent/auto-research works because LLMs now handle monitoring, analysis, and iterative optimization well enough to automate many research loops. SimGym’s value depends on historical data; without it, agents merely repeat prompts, but with it, they can be calibrated against real outcomes. SimGym models trajectories and counterfactual interventions, not just summary statistics, making it more powerful than naive A/B testing. Liquid models are best viewed as efficient, task-specific architectures for low-latency or long-context workloads, especially after distillation. Shopify’s AI strategy is not just model usage; it is a systems moat built from data, infra, and workflow redesign.

Data Points: AI tool adoption: Approaches 100% - Shopify’s internal daily active usage of at least one AI tool across the company December inflection: Major phase transition - Parkin said model quality improvements in December triggered rapid adoption growth PR merge growth: 30% month over month - A slide showed merge growth accelerating from roughly 10% to 30% Search throughput: 800 QPS to 4200 QPS - Parkin cited a Shopify search optimization project that increased throughput at the same quality SimGym target correlation: 0.7 correlation with add-to-cart events - Internal goal for simulation outputs to match real-world outcomes Auto-research experiment count: 400+ experiments - A hobby experiment ran for weeks and generated over 400 trials Successful experiments: 1 of 400 - Parkin described one long-running auto-research run with only one successful result Liquid search latency: 30 milliseconds end-to-end - A 300M-parameter Liquid model was used for search query understanding under strict latency constraints Liquid model size: 300 million parameters - Example production model used for search Offline model range: 7–8 billion parameters - Used for batch tasks like catalog classification and attribute extraction Shopify customer simulation scale: Decades of historical data - SimGym depends on long-term merchant/customer behavior history Previous Sydney role detail: Shipped first in India - Parkin said Sydney was first deployed in India and initially not widely noticed

Pivotal Quotes: "it approaches really a hundred percent by now it's hard not to do your job now without interacting deeply at least with one tool" — Michael Parkin: On internal AI adoption across Shopify "the anti-pattern is running multiple agents too many agents in parallel that don't communicate with each other that's almost useless" — Michael Parkin: On why token volume alone is not a good productivity metric "if you're not using auto research like approach in whatever you do then you're missing out" — Michael Parkin: On the value of Tangent-style autonomous experimentation

Implications: Shopify is turning AI into operating infrastructure, not just a feature. For the industry, the message is clear: proprietary data, reproducible workflows, and critique-heavy agent systems may become the real competitive moat.