Poolside engineers use the $3B coding model to automate software development. Jason Warner and Eiso Kant founded the company in May 2023. Warner previously served as the Chief Technology Officer at GitHub and Heroku. The organization manages 150 employees worldwide as of December 2025. The company reached a $3 billion valuation in October 2024 after a $500 million Series B round led by Bain Capital Ventures. This round included participation from DST Global, Citi, and Nvidia. Earlier, the company secured $126 million in October 2023 at a $475 million valuation led by Felicis Ventures. Total cumulative private capital exceeds $626 million. The company attracts researchers from DeepMind, Google Brain, and Meta AI. Poolside develops models that understand software architecture, testing patterns, and deployment practices. It uses Reinforcement Learning from Code Execution Feedback to train these models from scratch. This technical approach provides a distinctive capability for software engineering. The company’s mission aims to build artificial general intelligence through software engineering. During investor demonstrations, the model generated code for a snake video game in python using a natural-language prompt. Poolside occupies a position alongside peers like Cognition AI, Cursor (Anysphere), and Codeium/Windsurf. It also competes with frontier AI lab alternatives such as Anthropic Claude Code and Google DeepMind Jules.
Laguna model specifications
The Laguna family includes the S 2.1 and XS 2.1 models. Laguna S 2.1 contains 118 billion parameters with 8 billion active parameters and a 1 million context window. The Laguna XS 2.1 model offers 33 billion parameters with 3 billion active parameters and a 256,000 context window for agentic coding tasks that require speed and efficiency on local devices. The S 2.1 model provides frontier-class reasoning at a mid-size cost. Released in July 2026, these models prioritize quality, speed, and efficiency.
| Model Type | Total Parameters | Active Parameters | Context Window |
|---|---|---|---|
| Laguna S 2.1 | 118B | 8B | 1M |
| Laguna XS 2.1 | 33B | 3B | 256K |
The company uses a "Model Factory" to train, evaluate, and iterate on these foundation models. This system includes GPU-to-GPU weight transfers and automated architecture ablations. The XS 2.1 model is the lightest and fastest agentic coding model. It is small enough to run on-device. The S 2.1 model focuses on frontier-class reasoning. The company researches high-performance alternatives like NVIDIA’s NVLink C2C on Grace-based superchips to offload MLP activations to host memory. This method achieves a 6% to 13% throughput improvement over selective activation checkpointing. In April 2026, the company released Laguna M.1 and Laguna XS.2 alongside the runtime used to train agents.
Deployment and productivity for Shopify developers
Developers deploying Shopify apps must move code from local environments to hosting services. This process involves connecting the hosted app to Shopify through the Shopify CLI or the Dev Dashboard. The Shopify React Router template manages authentication, session management, and webhook handling. Developers must set environment variables such as SHOPIFY_API_KEY, SHOPIFY_API_SECRET, and SCOPES. The application URL must match the SHOPIFY_APP_URL environment variable in the shopify.app.toml configuration file.
Poolside supports enterprise customers through direct deployment and partnerships with the Department of Defense and Amazon Web Services. The company provides foundation models on Amazon Bedrock and Amazon Elastic Compute Cloud. The Paris branch supports the company’s global engineering and product leadership.
Developers using the Shopify React Router template often use Vite to build bundles for hosting. If the hosting provider supports Docker containers, the template includes a Dockerfile to build the app. For developers using SQLite, they must ensure the container provides filesystem access. Some providers, like Fly.io, might put web containers to sleep when idle, which resets the disk and wipes the SQLite database. In these cases, developers must create a separate volume for the database file to persist data across redeploys. To deploy, developers follow specific steps: create a new app configuration, set up a container and database, build and run the app, deploy the app configuration, and test the app. Developers must also use the Shopify CLI to manage app extensions and deploy new versions. This ensures that Shopify can reliably determine which extensions were added, updated, or removed.
Does the use of reinforcement learning from code execution improve the reliability of Shopify app extensions?
