The student writes the curriculum: agents are learning to improve themselves

The research campaign that ran itself

Four Claude Opus 5.5 agents ran a 104-hour post-training campaign on Qwen3-4B-Base with no human research intervention, working inside a sandbox and barred from using external teacher models. The agents managed data preparation, algorithm changes, and hyperparameter tuning themselves, and they were given no research advice along the way. Starting from a dev-suite score of 15.9, they reached 67.3; on a held-out suite they moved from 9.2 to 48.2. The vendor's own Qwen3-4B-Instruct-2507 scores 66.5 and 47.9 on those same suites — so the agent-trained model edged past the official instruct version on both. The agents invented no new algorithm; the paper says they recombined known methods — supervised fine-tuning on public reasoning data, reinforcement learning, weight merges, and self-distillation. The advance was the researcher, not the recipe. m:…3044

Terminal agents that rewrite their own attempts

Terminal-RSI takes the loop to the command line. The framework alternates recursive trajectory rewriting — the agent rewriting its own past attempts — with critic-inheriting policy refinement, so each iteration starts from a better critic. The resulting Terminal-RSI-27B scores 86.52% on Terminal-Bench 2.1 and 79.00% on Terminal-Bench Hard, but only 19.84% on the newest Terminal-Bench 4.0 and 16.95% on 3.0. On LHTB, its process reward rose from 0.21 to 0.37 and strict completion went from 0 of 46 to 5 of 46. The numbers are honest about how early this is: the system is getting better at grading and re-attempting its own work, but five strict completions out of 46 is a beginning, not a capability. m:…0341

When the scaffold is the student

The Self-Evolving Agent Harness paper pushes the idea further from the weights. Its framework uses feedback-driven revisions to evolve the harness itself — prompts, memory, and workflows — while keeping the base model's weights fixed, and accepted revisions persist across runs. The authors released the code as the Harness RSI repository and attach a theoretical generalization bound to the process. But they also warn that reliable deployment requires continual validation as task distributions and evaluators change: a harness that evolves against one evaluator can silently drift when the world moves. This is self-improvement with a maintenance contract attached. m:…7020

TimeRSI shows what the loop looks like in production-shaped work. Instead of retraining a frozen time-series foundation model, a coding agent iteratively revises the forecasting pipeline itself — translating task descriptions into executable changes, testing each revision on historical windows, and using observed errors to guide the next one. An adaptive selector then picks the best pipeline for deployment conditions. As of October 15, 2026, the approach pushed TimesFM 3.0 to number one on both the Fev Bench and TIME leaderboards, outperforming all submissions on six point and probabilistic metrics across 198 tasks. The model never changed; the system around it did. m:…2152

The decision: fund the loop, not the checkpoint

The pattern across all four papers is the same: the unit of improvement is moving from the weights to the scaffolding. For builders, this suggests the next durable investment is harness and evaluation infrastructure — the machinery that lets an agent attempt, grade, and revise — rather than chasing the next checkpoint release. The caveats are real: Athanor's result is one campaign on one base model with no research advice given, so reproducibility is unproven; self-evolving harnesses need the continual validation their authors prescribe; and TimeRSI's gains are in time-series forecasting, not everywhere. But the direction is unmistakable — the student is writing the curriculum now, and the curriculum is code. m:…3044 m:…7020