Descriptions:
Jason Warner and Eiso Kant, co-founders of AI startup Poolside, take the stage at the AI Engineer conference to demonstrate their Malibu Agent — a coding AI built entirely from scratch using reinforcement learning combined with next-token prediction, trained over two and a half years without relying on models from OpenAI or Anthropic.
The centerpiece of the session is a live demo set in a real-world context: converting a legacy ADA codebase (a programming language used in government satellites and defense-critical infrastructure) to Rust, entirely within VS Code. The Poolside agent autonomously writes 1,152 lines of Rust, runs tests, and iterates — completing the task quickly. Warner emphasizes a key design constraint for defense environments: agents must operate within strict permission boundaries and cannot freely access data sources or execute arbitrary commands, a requirement that directly shaped Poolside’s agent architecture.
The team also announces that their next-generation model — beyond the current Malibu Agent — will be released publicly in early 2026, available through Poolside’s own API and on Amazon Bedrock. This positions Poolside as a fifth frontier-level model provider alongside OpenAI, Anthropic, Google, and Meta, targeting both developer tool platforms like Cursor, Windsurf, and Cognition, as well as vertical AI applications such as Harvey and Writer.
📺 Source: AI Engineer · Published December 27, 2025
🏷️ Format: Keynote Launch
