← All organizations

AI coding models and enterprise agent systems

Poolside

Poolside offers Laguna, an open-weight family of agentic coding models, and tools for developers and enterprise teams. Its macOS Desktop Assistant runs multiple coding agents across projects and repositories, with extensions for VS Code and Visual Studio; the smaller Laguna XS model can run on-device. The Poolside Platform deploys agents inside enterprise security boundaries with auditability and governance. Its 2025 acquisition of Fern Labs added Bridge, a multi-agent orchestration layer, and engineers who work directly with customers on deployments.

Founded in San Francisco in 2023, Poolside is led by co-founders and co-CEOs Jason Warner and Eiso Kant. Its research uses reinforcement learning from code execution: compilation results and tests give models automated feedback as they practice software tasks. Its Model Factory automates training, evaluation and experiments with model architectures and data, making those experiments reproducible. These systems underpinned its foundation-model development before its reported strategic transition in 2026.

In August 2026, Nvidia reportedly agreed to a $6 billion non-exclusive licence for the Model Factory, job offers to 109 employees, and a separate $1 billion investment at a $12 billion pre-money valuation. Poolside remains independent, with its founders staying; the arrangement is not an acquisition. Poolside reportedly stopped building frontier models and had not disclosed its updated vision.

poolside.ai

Start here

  1. Your agent is blindfolded

    Start here to understand why Lajili treats an agent’s ability to inspect applications as more consequential for reliability than the greenfield–brownfield distinction.

    Johan LajiliAI Engineer Europe 2026

  2. The Messy Reality of Scale: Synthetic Data and Pre-Training

    Learn how synthetic-data generation fits into Laguna pre-training, and how hardware faults and tensor-parallel accumulation can undermine a training run.

    Marah Abdin · Robert McHardyAI Engineer World's Fair 2026

  3. AGI: The Path Forward

    Use this talk for the broader rationale behind pursuing AGI through software-engineering models trained with reinforcement learning.

    Eiso Kant · Jason WarnerAI Engineer Code 2025

Messages from the stage

From coding models to deployment infrastructure

Jason Warner and Eiso Kant pair an Ada and Rust agent demonstration in Visual Studio Code with discussion of independent model serving, Amazon Bedrock distribution, and plans for large-scale compute.

Data quality meets numerical stability

Marah Abdin describes rephrasing, code and STEM pipelines, orchestration, and validation. Robert McHardy examines how defective GPUs, silent data corruption, and insufficient BF16 precision complicate large-scale training.

Give agents an application interface

Johan Lajili describes Spoolside, an internal CLI supplying screenshots, compressed application snapshots, and interaction capabilities. He recommends product-specific interfaces through CLIs, skills, or MCP integrations.

Affiliations reflect each recorded session, not necessarily current employment.

Company sources · checked 2026-08-28