Radar, 17 Oct 2026

AutoEnvScaling: Automating the Data Flywheel with Terminal Agents

The loop: A proposer agent builds new reinforcement learning environments from a solver model's rollouts. These environments are used to train the solver, which improves its performance on terminal tasks and generates better rollouts for the next round of environment design.

A framework that treats environment design as a terminal task, using agents to automatically generate and scale verifiable training environments for reinforcement learning.

MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering

The loop: An agentic model uses a model-directed interaction harness to autonomously organize machine learning experiments and adapt optimization strategies. The resulting long-horizon interaction trajectories provide supervision that improves the model's own machine learning engineering capabilities.

A post-training method and agentic model designed to autonomously manage long-horizon machine learning experimentation and learn from its own execution feedback.

Reels-RSI: Program-Level Self-Improvement for Code-Rendered Explainer Videos

The loop: A reel director agent logs its code revisions and pairs check findings with the diffs that fixed them. It compiles these retrospectives into unit-tested rules that update its playbook, improving its ability to generate code-rendered explainer videos.

A self-improving reel director that enhances its video generation capabilities by learning program-level rules from its own code revisions and static check findings.

A Simpler and Better Debate Protocol for Scalable Oversight

The loop: AI agents debate complex tasks to decompose them into simpler claims that can be easily judged. This process generates reliable verifiable rewards that serve as the training signal to improve the capabilities of future models.

A theoretical framework for an AI debate protocol that provides worst-case correctness guarantees for decomposing complex problems to generate verifiable training rewards.

Myelinate

The loop: An AI agent framework autonomously acquires and retains new skills and memory during operation. These additions improve the agent's capability to execute future tasks within the same framework.

An open-source repository providing self-improving skills and memory capabilities with built-in governance for AI agent frameworks.

Sergey Brin says Google's AI models are reaching recursive self-improvement

The loop: Google's AI models perform work that improves their own architecture or training, which makes the successor models better at that work.