Poolside trains its own models with reinforcement learning for coding and other long-horizon knowledge work.
2
Its public demo shows an agent converting Ada code to Rust, testing the result, and adding a command-history feature inside Visual Studio Code.
3
Poolside plans to release its next-generation model through its own API and Amazon Bedrock, while scaling training with more than 40,000 GB300 GPUs.
Summary
Jason Warner and Eiso Kant describe Poolside's plan to build models for software development and wider knowledge work. They argue that next-token prediction needs reinforcement learning to produce more capable agents, and say Poolside has trained two generations of models from scratch. The live demo uses Visual Studio Code and an Ada codebase from critical infrastructure. Poolside Agent explains the code, converts it to Rust, writes tests, runs the program, and adds command-history support after a follow-up request. Warner also discusses permission controls for high-consequence environments, where agents cannot be given unrestricted access. Kant says compute was the main missing ingredient and points to more than 40,000 GB300s coming online. The company plans a public model release through its own API and Amazon Bedrock. Both speakers expect agent interfaces to change as tasks stretch from hours toward days, while inviting companies to build on Poolside models.
Poolside pairs next-token prediction with reinforcement learning
Jason Warner says Poolside was founded to close the gap between models and human intelligence. The company trains its own models from scratch and began with the view that next-token prediction was a major technical advance, but needed reinforcement learning to make a larger jump in capability. After two and a half years, Poolside is on its second generation, called Malibu Agent. The company is applying this work to coding first, while also aiming at other forms of knowledge work. Warner presents the approach as a long-term effort to build more capable models rather than as a wrapper around another provider's system.
The demo uses an agent to work on unfamiliar Ada code
The demonstration starts with an Ada codebase used for critical infrastructure. Eiso Kant asks what the codebase is about, and Poolside Agent explains it inside Visual Studio Code. Warner says the model was trained from scratch using Poolside's own techniques. The agent then receives a request to convert the database to Rust. It creates about 1,152 lines of code, produces test commands, and shows file changes in a live diff pane. The speakers position this as an example of the coding workflow Poolside is building, while acknowledging that the interface could change over the next few years.
Warner says Poolside has spent the past year working inside government and defense environments, where code can have serious consequences. In those settings, an agent cannot simply run around and access data sources without limits. It needs the right permissions, and its actions must be restricted to what an organization is willing to allow. This concern shapes the product beyond the coding model itself. Poolside describes its stack as vertically built, with its own model and interfaces, and says the company is working in environments where unrestricted autonomy is unacceptable.
The agent can test generated code and respond to follow-up requests
After generating the Rust conversion, the agent explains how to run both the Ada and Rust versions. The presenters run the Rust program, inspect warnings, create a table, insert a record, and query the result. They then ask for command-history support through the up arrow. The agent identifies the Rustyline library, updates the relevant files, builds the project, and writes scripts to test the history feature. The demo ends with the up-arrow command working. The workflow shows the agent carrying a request through code changes, builds, tests, and a second feature request rather than stopping after producing a code sample.
Poolside is building a general knowledge-work platform around its models
Warner says Poolside's models are not limited to coding agents. He describes them as useful across knowledge work and mentions other ways to interact with the platform, including a web interface and an agent that can be downloaded to a machine. The coding experience shown in Visual Studio Code is one interface into the system. He jokes that the models can also write poems and love letters, but says Poolside does not promote those uses. The company's stated focus remains models that can help with engineering and other work inside organizations.
The next model is planned for public access through Poolside's API and Bedrock
Warner says Malibu Agent is Poolside's second generation and that more compute is coming online for the next generation. He expects that model to become publicly available early the following year, behind Poolside's own API and on Amazon through the Bedrock API. He names coding products such as Cursor, Windsurf, Cognition, and Replit as possible users, along with applications such as Harvey and Writer. Poolside wants developers to consume the model directly rather than limiting it to the company's own interface. The release plan is presented as a way to let other products build on Poolside's model capabilities.
Compute is the main resource Poolside says it was missing
Kant describes the company's progress after two and a half years of starting from scratch. He says Poolside's models have become increasingly intelligent, but that compute was the missing ingredient. More than 40,000 GB300s are coming online, which he says will let the company scale models further for software development and other long-horizon knowledge work. Poolside is building data centers at a multi-gigawatt campus in West Texas and describes its work as vertically integrated from infrastructure through models. Kant says the current interface is only one expression of the intelligence they are developing.
Kant says both he and Warner already run agents that work asynchronously on tasks for hours. He expects agents to handle tasks lasting days in the coming years. As the duration and scope of work change, he expects the user interface to change as well. Poolside is concentrating on model intelligence and the ability to scale and serve it, while Kant expects the conference audience to develop better ways to turn that intelligence into economically valuable work. He frames the current period as an awkward stage before AGI, where companies are still learning how to connect increasingly capable models with real-world problems.
"You need to have the right permissions. You got to actually really ratchet these things down to do things inside those environments that they feel comfortable with."03:44
Who should watch
You are evaluating coding agents for regulated, government, or defense environments and need to understand how Poolside approaches permissions.
You are building an engineering or knowledge-work product and want to use a model through an external API or Bedrock.
You are interested in long-running agents and the infrastructure needed to train and serve them.
