Building Autonomous ML Experimentation with Tangle and Tangent

Shopify has released Tangle and Tangent, an open-source ecosystem designed to automate machine learning (ML) experimentation. According to a report from Linux.com, Tangle provides a platform-agnostic pipeline execution environment, while Tangent acts as an autonomous agent that orchestrates those workflows by forming hypotheses, running experiments, and analyzing metrics.

The system aims to replace the manual “loop” of ML engineering—building pipelines, reading metrics, and adjusting parameters—with an automated process. Shopify engineers developed the tools to allow ML experiments to run on a shared platform where results are reproducible and pipelines can be modified without cloning private notebooks.

Tangle’s Pipeline Architecture and Caching

Tangle is an open-source ML experimentation platform featuring a drag-and-drop visual editor. Users create pipeline graphs by wiring outputs to inputs and submitting them for execution in the cloud or locally, according to Shopify developers.

A core technical feature of Tangle is its caching layer. This system skips or reuses previously executed steps, including those currently in flight, to reduce the cost and time of iteration. Shopify states that all pipeline runs, including logs and components, are stored indefinitely to ensure reproducibility over several years.

The platform is designed for interoperability. Any containerized CLI program written in any language can function as a Tangle component. These components exchange data using standard file formats such as JSON, Parquet, and CSV.

Tangent: Autonomous Agent Orchestration

Tangent is an autonomous agent that operates on top of Tangle to accelerate experimentation. According to Shopify, Tangent follows an “autoresearch” pattern, moving beyond single training scripts to manage full experiment pipelines using a fleet of specialized subagents.

The agent operates via an eight-step loop: initialize, analyze, hypothesize, submit, monitor, evaluate, synthesize, and decide. To prevent “drift” during long runs, Tangent uses gated checkpoints. The agent cannot advance to the next step until every item on a specific checklist is passed, at which point it reloads its instructions and context.

Tangent’s “brain” is defined by skills written in Markdown. This approach allows skills to be portable and reviewable via pull requests without requiring a binary build or a proprietary client. The system utilizes several subagent roles, including a researcher, builder, debugger, and reviewer.

Security and the Agent Hosting Platform

To support these agents, Shopify built a Linux-based hosting platform. Each Tangent instance runs as a Linux-based VM or container, which allows it to use standard Linux networking and storage primitives. In Kubernetes deployments, each instance is a StatefulSet with a per-instance PersistentVolume.

To prevent agents from leaking credentials to AI providers, the platform uses an Auth Proxy. This proxy lives in a separate container and intercepts HTTP requests to automatically add authentication headers. For HTTPS requests, the proxy generates SSL certificates on the fly, which the agent container trusts via a generated certificate authority.

The Tangent Shell provides a remote environment where agents maintain memory and sessions across restarts. This shell is open source and can be extended through Agent Bundles, which are packaged sets of prompts, tools, and workflows.

Reranking Model Case Study

Shopify applied Tangent to rebuild a large reranking model. An engineer provided the direction regarding features and training data, while Tangent executed the loop of building and analyzing experiments.

The results showed incremental improvements in precision across different recall (R) metrics. According to the provided data, the baseline pipeline started with 67.3% precision at R@90%. After migrating to a reproducible trainer, this rose to 69.5%. The most significant lift occurred after adding richer product features (structured metadata and taxonomy), which pushed R@90% precision to 71.3%. The final addition of search-derived and hard-negative training data resulted in a peak R@90% precision of 75.6%.

Tangle, the Tangent skills, the Hosting Platform, and Tangent Shell are all released under the Apache 2.0 license. The projects are maintained by Alexey Volkov, with Shopify serving as the initial sponsor and infrastructure steward.