Today's edition highlights the growing trend of AI agents modifying their own runtime environments and harnesses, moving beyond simple prompt tuning to structural self-improvement.
The loop: An AI agent modifies its own harness, prompts, and meta agent to solve and verify open scientific problems. These modifications pass through a kernel-enforced gate and are directly used by the agent to conduct further autonomous research.
A self-improving research agent that autonomously modifies its own harness to solve and verify open problems in quantum information.
The loop: An agent searches for and curates training data to post-train its own successor across multiple domains. The resulting model is evaluated on generalization benchmarks, and the results guide the agent in the next round of data selection.
A multi-domain benchmark for evaluating recursive self-improvement where an agent searches for and curates data to post-train its successor.
The loop: An offline evaluation harness drives GEPA-based prompt optimization for a production conversational agent. The resulting compressed prompt is deployed back to the agent, shrinking its prompt surface while maintaining or improving task correctness.
A reproducible offline evaluation harness enables safe, automated prompt compression for production conversational agents without regressing performance.
The loop: A foundation model synthesizes its own curriculum of queries and privileged reasoning traces guided only by high-level topic descriptions. These synthetic traces are used for on-policy self-distillation, improving the model's performance across multiple domains.
A data-free self-training framework where a foundation model generates its own queries and reasoning traces for self-distillation.
The loop: Personal agents evolve their own harnesses to optimize performance on tasks. The evolved harness is then deployed as the runtime environment for the agents, improving their execution of future tasks.
A study on personal agents that treat the evolution of their own harnesses as a learning mechanism to optimize future task execution.
The loop: An autonomous system automates large-scale omnidisciplinary research tasks using a prompt economy. This automation improves the efficiency of the research process, which is then used to develop better AI systems.
An autonomous research system designed to automate large-scale omnidisciplinary research tasks.
The loop: AI models analyze telemetry to optimize GPU yield and cluster resiliency in AI factories. This optimization improves the hardware infrastructure that trains and runs those same AI models.
NVIDIA uses AI to analyze telemetry and optimize GPU yield in large-scale AI factories.
The loop: Codex Ultrafast helps engineers investigate incidents and refactor code. This accelerates the development and security analysis of the underlying AI systems, including Codex itself.
OpenAI engineers use Codex Ultrafast to accelerate incident investigation and code refactoring.
