Poolside, the AI lab betting everything on software development

A company built on one bet: code first

Poolside was co-founded by Jason Warner, former CTO of GitHub and now its CEO, and Eiso Kant, its CTO. Its official pages do not give a founding year or a headquarters city; press coverage describes it as founded in 2023, with teams in the United States and France.

Its thesis is stated plainly: according to Poolside, the fastest path to artificial general intelligence runs through mastering software development, especially long, multi-step tasks. The whole company is therefore geared toward agentic coding: models that do more than complete a line, reading a repository, running commands and fixing their own mistakes.

The Laguna family

As of October 22, 2026, Poolside features the Laguna family, released as open weights:

• Laguna XS 2.1, the lightest and fastest model, built to run locally;
• Laguna S 2.1, larger, designed for long-horizon tasks and heavier use of reasoning.

Both use a mixture-of-experts architecture: only a small share of the parameters is active for each generated word, which cuts compute costs. Earlier 2026 releases (Laguna XS.2 and Laguna M.1) paved the way, with Laguna XS.2 published under the Apache 2.0 license. Poolside says licenses vary by model, from OpenMDW to Apache 2.0, all of them permissive.

How to access the models

Poolside offers several entry points, described in its official documentation:

• weights on Hugging Face, usable with vLLM, SGLang, Transformers or Ollama;
• an OpenAI-compatible API;
• pool, a coding agent that runs in the terminal;
• intermediaries such as OpenRouter and Vercel AI Gateway.

For businesses, Poolside Platform deploys agents inside the customer’s own security perimeter, and a macOS app lets developers run several coding agents in parallel.

What sets Poolside apart

In practice, Poolside stands out through specialization: where other labs aim for a do-everything assistant, it focuses research and products on software. The company also publishes a series of technical posts detailing its training pipeline, which is fairly rare. To compare Laguna with other open models, see our page on open-source models and the model comparison tool.