AGI: The Path Forward – Jason Warner & Eiso Kant, Poolside

2025-12-27 · Source: AI Engineer · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Intermediate, long

Summary

Poolside, a company founded 2.5 years ago, is developing its own AI models, currently in their second generation (Malibu agent), to bridge the gap between models and human intelligence. Their approach combines next-token prediction with reinforcement learning. The company demonstrated its Poolside agent within Visual Studio Code, showcasing its ability to convert ADA code, a language used in critical government infrastructure, into Rust. This conversion involved generating approximately 1152 lines of code, followed by automated testing and debugging. Poolside operates in high-consequence environments, primarily within the government and defense sectors, emphasizing the need for agents with strict permissions. The demo also featured the agent adding a new feature (up-arrow command history) to the converted Rust application. Poolside plans to release its next-generation model publicly early next year via its own API and Amazon Bedrock, aiming to support both engineering assistants and other AI applications.

Key takeaway

For CTOs and VPs of Engineering managing critical infrastructure or defense projects, Poolside's approach to AI-driven code conversion and feature addition in high-consequence environments offers a compelling solution. Your teams could leverage their Malibu agent to modernize legacy codebases like ADA to Rust, potentially reducing technical debt and improving maintainability, while ensuring the necessary security and permission controls are in place. Consider evaluating their upcoming public API release for integrating advanced, secure coding agents into your development workflows.

Key insights

Poolside combines next-token prediction with reinforcement learning to build highly capable AI agents for high-consequence environments.

Principles

• Reinforcement learning enhances next-token prediction.
• AI agents require strict permissions in sensitive environments.

Method

Poolside's agent converts legacy ADA code to Rust, including automated testing and debugging, and can add new features, all within a VS Code interface, using proprietary models trained from scratch.

In practice

• Convert legacy codebases (e.g., ADA to Rust).
• Automate code testing and debugging.
• Integrate AI agents into existing IDEs like VS Code.

Topics

• Reinforcement Learning
• Large Language Models
• Code Generation
• AI Agents
• High-Consequence Environments

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, Software Engineer

Related on AIssential

• Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment — AI systems are demonstrating advanced capabilities in complex coding and face diverse attack vectors as they become autonomous agents.
• From Dashboards to Decisions: How We’re Building AI Agents at RaySuite AI — AI agents transform user interaction by providing intelligent, contextual answers and taking action over complex systems.
• From idea to AI app: Creating intelligent research assistants with Strands — Strands Agents and Kiro simplify AI agent development, enabling complex applications with minimal code and LLM-driven reasoning.
• Ryan Carson Is a One-Person Code Factory — AI agents enable single founders to operate complex startups by automating engineering and operational tasks through iterative loops.
• not much happened today — AI agent tooling, infrastructure, and robotics are rapidly advancing, while platform stability and search access pose new challenges.
• The Verification Horizon: No Silver Bullet for Coding Agent Rewards — Reliably verifying coding agent solutions is harder than generating them, requiring co-evolving verification with agent capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.