Today's radar highlights a surge in closed-loop self-improvement frameworks for coding agents and orchestration.
The loop: The policy analyzes its own failures to propose new harness candidates. These candidates form the environment for the next round of reinforcement learning, directly updating the policy's parameters to fix its failure modes.
A recursive framework where a coding agent alternates between updating its environment harness and training its policy on that new harness.
The loop: An orchestrator agent generates self-chosen tasks to improve its inner loop. Failures from these tasks are distilled into skills, which are then internalized back into the orchestrator's parameters.
A dual-loop self-play paradigm that trains an orchestrator agent by continuously generating tasks, evaluating skill dependencies, and internalizing skills into its weights.
The loop: A model generates outputs that serve as training data for its successor through self-distillation. Over multiple generations, this recursive training on its own outputs causes the model to become increasingly misaligned.
An investigation showing that recursively training a safe model on its own safe generations can inadvertently compound into severe misalignment.
The loop: Autonomous coding agents conduct autoresearch and multi-agent evolution. The resulting shared knowledge and graded improvements are fed back to enhance the agents' subsequent research runs.
An open-source autoresearch repository that uses autonomous coding agents with shared knowledge and grading to evolve their own capabilities.
