Shopify has open-sourced Tangle and Tangent, two tools designed to automate and streamline machine learning experimentation. Tangle serves as a platform-agnostic, visual pipeline orchestration tool, while Tangent acts as an autonomous agent that manages the iterative process of hypothesis generation, pipeline execution, and performance analysis. Both projects are available under the Apache 2.0 license.
Tangle: Visualizing ML Pipelines
Tangle is an open-source experimentation platform that allows engineers to build and execute machine learning pipelines through a drag-and-drop visual interface. According to Shopify’s engineering team, the platform is designed to solve the friction common in iterative ML research, where environment mismatches and manual tracking often impede progress.
The system utilizes a caching layer that reuses previously executed steps, including those currently in progress, to accelerate iteration cycles. Because Tangle stores all pipeline graphs, components, and execution logs, it ensures reproducibility across the team. Any containerized command-line interface (CLI) program can function as a Tangle component, allowing engineers to exchange data in various formats such as CSV, Parquet, or JSON. The architecture functions similarly to Unix pipes, enabling modularity in pipeline construction.
Tangent: Autonomous Experimentation Agents
Tangent is an autonomous engineering agent built to run on top of Tangle. It automates the "autoresearch" loop—a concept popularized by Andrej Karpathy—by managing full experiment pipelines rather than individual scripts.
The agent operates through an eight-step loop:
- Initialize
- Analyze
- Hypothesize
- Submit
- Monitor
- Evaluate
- Synthesize
- Decide
Tangent utilizes "skills" written in Markdown to define its behavior. These skills are modular and portable, allowing for different sub-agents—such as a researcher, builder, or debugger—to handle specific tasks. The agent maintains persistent memory through plain-text files, ensuring that it retains lessons learned from previous iterations.
Security and Hosting Infrastructure
To address the risks associated with autonomous agents interacting with sensitive services, Shopify developed a specialized hosting platform for Tangent. This infrastructure uses a system-wide proxy to intercept and manage HTTP requests, injecting authentication headers without exposing credentials to the agent itself.
The hosting platform runs on Linux-based containers, leveraging standard networking and storage primitives. By deploying instances as Kubernetes StatefulSets, the platform ensures that agent sessions and persistent memory are maintained across restarts. This configuration allows for "agent bundles"—packaged sets of prompts, tools, and workflows—to be deployed without modifying the core codebase.
Performance and Practical Application
Shopify reported using Tangent to rebuild a large-scale reranking model. By automating the experimentation loop, the agent attempted various feature sets and training data configurations while measuring performance against defined metrics.
| Pipeline Stage | R@90% Prec. | R@95% Prec. | R@97% Prec. |
|---|---|---|---|
| Previous Pipeline | 67.3% | 54.4% | 33.6% |
| Standardized Pipeline | 69.5% | 51.9% | 35.2% |
| Richer Product Features | 71.3% | 58.7% | 48.5% |
| Expanded Training Data | 75.6% | 60.2% | 43.9% |
The results indicated that the agent could successfully iterate on the model, achieving performance gains through systematic testing. The projects are currently maintained on GitHub, where the team invites contributions for new sub-agent skills and system improvements.