Poolside AI

Poolside AI builds foundation models for agentic coding and long-horizon software work, treating software development as a path toward artificial general intelligence. Its research combines language-model training with reinforcement learning from executing code. In 2026, it began publicly releasing its Laguna models and expanded its developer tools from the terminal-based pool agent to a desktop workspace that can also run other companies’ agents.[4][6][9]

The public Laguna releases followed several years of enterprise-focused model development. Poolside’s founding team also brought earlier experience in developer tools and machine learning on code.[3][6]

Before Poolside, Eiso Kant founded source{d}. In his Latent Space interview, published July 23, 2026, he recalled pivoting his startup after reading Andrej Karpathy’s 2015 article on recurrent neural networks. The team pursued language models on code using RNNs, LSTMs and later Transformers through the end of 2019. Kant described the effort as a failure, involving about $12 million and roughly 40 people; he believed they had missed the importance of continuing to scale the models. These were source{d}’s expenditures and team, before Poolside.[1]

Jason Warner and Kant first met in 2017. Warner had led engineering at Heroku and Canonical and served as GitHub’s CTO, where he incubated Copilot. Poolside dates its founding in San Francisco to April 2023. Its founders chose coding as the first major capability to pursue, believing reinforcement learning would become a central axis for improving model intelligence.[3][4]

Kant described Poolside as an American company from its beginning, with a deliberate early decision to recruit researchers outside the Bay Area. That produced a globally remote organization, with offices including Paris and London. He also recalled an approximately five-person team spending three weeks resolving an optimizer bug in its training code, which it had written from scratch. He did not remember the exact bug, but regarded solving the instability as an important moment in the team’s confidence and engineering culture.[1]

On October 2, 2024, Poolside announced a closed $500 million Series B led by Bain Capital Ventures. NVIDIA, Citi Ventures, Capital One Ventures and other corporate and institutional investors participated. Earlier financing comprised a $26 million seed round led by Redpoint Ventures and a $100 million Series A led by Felicis Ventures. Poolside reported that the capital had brought 10,000 NVIDIA GPUs online, while also funding its commercial work and research organization.[3]

Poolside’s core approach, Reinforcement Learning from Code Execution Feedback (RLCEF), gives models feedback on attempted software solutions. They write and execute code, check the result and try again. Compilation results, unit tests and performance benchmarks supply automated signals, making software a setting where models can learn through repeated experiments without physical-world interaction. In October 2024, Poolside reported models exploring millions of tasks across 130,000 real-world codebases.[2][4]

In its December 4, 2024 AWS partnership announcement, Poolside said customers could contract its products and models directly from AWS and run them within their AWS Virtual Private Cloud. Its then-named malibu and point models supported inference on Trainium or NVIDIA hardware. Poolside distinguished customer-specific fine-tuning on private code and data from training its foundational models, for which it said it did not use customer data. It also described support for air-gapped, on-premises and other cloud deployments.[11]

On November 18, 2025, Poolside announced its acquisition of Fern Labs, the London company behind Bridge, a multi-agent orchestration layer. Fern’s founders—Ash Edwards, Alex Goddijn and Taylor Young—had previously led complex deployments at Palantir. The acquisition combined Poolside’s models and research infrastructure with Bridge and engineers embedded with customers to design and deploy agents within private enterprise environments. Poolside nevertheless acknowledged that reliable multi-agent operation in high-stakes settings remained an open problem.[10]

Poolside built the Model Factory as an internal framework joining data processing, pre-training, inference, evaluation and reinforcement learning. Joe Rowell and Eiso Kant’s July 2025 engineering account describes a Kubernetes orchestrator scheduling dependent workloads across a reported 10,000-H200 cluster. Evaluations can run during training; modular reinforcement-learning components let researchers introduce agents and rewards without rebuilding the surrounding pipeline.[5]

A consequential design choice was to stream data into training instead of waiting for a single enormous dataset to be fully materialized and distributed. Researchers can change blends and quantities across experiments or during training. Versioned experiment configurations and data lineage make results reproducible: the described stack uses Spark for data experiments, Iceberg for versioned assets and Dagster for workflows. Titan, its PyTorch-based training library, is closely integrated with Atlas inference so the Factory can check that training and serving implementations agree.[5]

The July 2025 account reported a code execution environment containing around a million repositories and many of their revisions, built into isolated containers with executable test suites. By his July 2026 interview, Kant estimated roughly 10,000–20,000 experiments per month, while noting that he had not checked the latest count. He described fewer than 70 researchers and another 35 engineers, with agents increasingly writing code, launching jobs and evaluating results under human direction.[5][1]

Poolside’s research agenda also extends beyond generating new experience through coding tasks. Its July 2025 statement, co-authored by Pengming Wang, Nikolay Zinov, Kant and Warner, argues that web text records finished answers while omitting much of the thinking that produced them. The company proposes reinforcement-learning exploration as a way to recover more reasoning from that existing material. This remains a research belief: the July 2026 S 2.1 release explicitly distinguished its demonstrated coding work from the web “decompression” approach still under development.[4][8]

Kant said he and Warner resumed discussions about openness at the beginning of 2026. His rationale was a preference for a world with “100 foundation model companies” over one with five, even if Poolside were among the five. He distinguished publishing model weights from sharing research that helps others recreate the work, while acknowledging unresolved business-model questions and the possibility that future capabilities could require release restrictions.[1]

On April 28, 2026, Poolside made its first public model shipment: Laguna M.1, with 225 billion total parameters and 23 billion active, and Laguna XS.2, with 33 billion total and 3 billion active. M.1 had been pre-trained in 2025; XS.2 was a newer generation, trained from scratch and released fully post-trained about five weeks after pre-training began. XS.2’s weights carried an Apache 2.0 license. Poolside credited approximately 60 people across its Applied Research organization for the models and introduced pool and Shimmer, a cloud development environment for web apps, APIs and CLIs, into preview.[6]

The July 2 release of Laguna XS 2.1 retained XS.2’s 33B-total/3B-active mixture-of-experts architecture and changed its license to OpenMDW-1.1. Poolside added quantized checkpoints and DFlash speculator models for local inference, and served it at 256K context through its API and OpenRouter. It announced plans to retire XS.2 from its API after one week while keeping that model available through Baseten for dedicated deployments.[7]

Laguna S 2.1, released July 21, scaled the family to 118 billion total parameters with 8 billion active per token, up to one million tokens of context, and thinking and no-thinking modes. Pre-training began May 22 on 4,096 NVIDIA H200 GPUs. It used the same pre-training data as XS 2.1, with scale, code fixes and recipe changes; post-training combined supervised fine-tuning partly using synthetic data with reinforcement learning on harder tasks. Changes included FP8 reinforcement learning, longer rollout budgets, new sandbox infrastructure and rollouts across multiple harnesses. Its weights were released under OpenMDW-1.1.[8]

Poolside reported 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1 with thinking enabled in its pool harness. These evaluations used an internal Harbor fork, repeated attempts per task and extended timeouts of five or six hours. DeepSWE comparisons involved differing harnesses, which Poolside acknowledged made scores less comparable. The release also identified limitations in third-party tool-schema adherence, nested JSON tool arguments and excessive thinking duration.[8]

An illustrative company case study put S 2.1 in an automated loop optimizing Poolside’s own harness: benchmark each change and retain measurable improvements. The model replaced quadratic string concatenation with buffers and reduced redundant copying and allocation. Poolside reported about 5.2% faster harness benchmarks and roughly 70% lower memory allocation, qualified as benchmarks rather than a full production test; the final artifact passed checks including Go’s race detector and go vet.[8]

On July 28, 2026, Poolside released Desktop Assistant for macOS, alongside Assistant extensions for VS Code and Visual Studio. Product Engineering Lead Andy Appleton described a design separating the interface, agent harness and model through the Agent Client Protocol (ACP). The default paired pool with Laguna S 2.1, while compatible harnesses such as Claude Code and Codex could run alongside it. pool could also connect to OpenAI-compatible inference endpoints.[9]

Assistant supports parallel sessions across projects, conversation handoffs with context, and Git worktrees that isolate concurrent tasks in separate checkouts and branches. Local model support uses MLX, including an offline setup with Laguna XS 2.1. Its model-fit estimates address a practical constraint Appleton emphasized: local coding models require substantial unified memory, and successfully loading a model does not necessarily make it comfortable for daily work.[9]