AI agents enter autoresearch era—but is it real progress?

The September-to-October 2026 wave mattered because it built on working precursors, not just theory. On 2025-05-14, Google DeepMind announced AlphaEvolve, an LLM-based coding agent that generated code changes, ran automated evaluators, and iteratively improved algorithms through evolution. By early March 2026, Karpathy’s open-source autoresearch turned that pattern into a practical loop with a measurable verifier and strict train-eval separation, where the model plans, acts, checks its own result, then repeats without human back-and-forth.

MarkTechPost says Karpathy pointed it at his already-optimized nanochat GPT-2 training code, and it ran for two days and completed about 700 experiments, keeping 20 genuine improvements. Those fixes cut GPT-2-quality training time by 11%, from 2.02 to 1.80 hours, and the loop was concrete: the agent reads the code, proposes a change, trains five minutes, then keeps or rolls back. Shopify’s Tobi Lütke reported a 19% improvement after 37 experiments, and by September and October 2026 multiple groups were showing self-evolving systems, including Bilevel Autoresearch, Dream-RSI, and co-evolving agent frameworks.

What changed in late 2026 was not just better scoring, but a new training logic: agents began creating the next problems they should learn from. In the AI Engineer demo, an agent launches training jobs, runs an offline evaluation framework, reviews production traces, adds new tasks to “hill climb against,” and then “try to write a new variant for itself,” while Daily Paper Cast described an online experiential loop that is “reward free,” using only textual feedback from interaction trajectories to extract knowledge, consolidate it into parameters, redeploy, and repeat.

That matters because recursive autoresearch turns self-improvement from a hand-labeled supervision problem into an agent capability. 每日AI论文’s analysis of AREX-2 frames that capability as iteratively refining solutions at test time through reflection plus long-horizon execution, while Hugging Face Daily Papers reported Athanor’s 104-hour campaign in which four Claude Opus 5.5 agents, given no research advice, pushed Qwen3-4B-Base from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on a held-out suite, edging the official supervised instruct baseline of 66.5 and 47.9. The same pattern appears even when the task itself is not prewritten by humans. Hugging Face Daily Papers described LongChang as a free-exploration system in which agents investigate their own questions, use exploration records to guide scene evolution without human intervention, and, crucially, showed that this loop can produce real discovery: among 12 LLMs, GPT-6-Astra achieves a discovery rate of 16.7%, while the other 11 score 0%, underscoring that autonomous generation, evaluation, and refinement can function as a genuine self-optimization mechanism rather than a scripted benchmark routine.

These systems improve by letting agents generate tasks, test outputs, reflect on failures, and feed the resulting experience back into search, policy updates, or model weights without relying on human-labeled rewards. Into AI reported that Meta’s Dr. Zero uses a proposer-solver pairing in which generated questions train the solver, and solver performance pushes the proposer toward harder but still learnable tasks, while Alibaba’s RM-NLHF turns model-generated critiques into process rewards and updates a Meta Reward Model online during training rather than depending on fresh human labels.

The same pattern appears in systems that learn from interaction traces instead of hand-scored supervision. Into AI described ERL as an explicit experience-reflection-consolidation cycle where a model makes an attempt, gets environmental feedback, reflects, and then internalizes successful revisions through RL and self-distillation, with gains of up to 81% in complex multi-step environments and up to 11% in tool-using reasoning tasks, plus 26.2% and 116.2% relative improvements in overall accuracy on the General AI Assistants benchmark and Humanity's Last Exam, respectively. Daily Paper Cast showed the loop can even build its own training substrate, where the search agent finds relevant information, the code agent generates the environment code, and the test agent validates the environment, then synthesizes 1622 conversations for supervised fine tuning and feeds verified trajectories into later updates.

The clearest evidence is that looped systems are no longer just promising better training economics; they are posting concrete deltas against older workflows. Matthew Berman described Andre Karpathy’s open-source loop as one that “gets run 24 hours a day” and acts as “truly an automated AI researcher,” autonomously designing and running architecture experiments that found improvements Karpathy himself had missed, while Brain Corp said BrainOS Clean 2.0 with SelfPath AI “eliminates the human training phase entirely,” replacing teach-and-repeat with robots that autonomously map facilities, choose productive paths, and adapt when layouts change.

Those gains also show up in benchmarked reasoning and deployment metrics. Brain Corp tied the removal of that labor bottleneck to scale—50,000 units worldwide and 5.3 million autonomous operating hours in the first half of 2026—while BDH-CQ, built by Pathway, Bielik AI, and NYU, reached 29.5% pass@2 on ARC-AGI-1 for $0.0007 per task, about 57 times cheaper than GPT 5.6 Luna (Low); Alibaba said Qwen3.8-Max outperforms prior Qwen models on agentic coding and long-horizon work, and NVIDIA’s Cosmos 3 reported measurable benchmark performance across multiple physical-AI-relevant evaluations, ranking number 1 across all five listed tests as evidence of improved reasoning and world-modeling quality. Cosmos 3 also includes Super, Nano, and Edge models built for simulation, reasoning, and edge deployment workloads, with Cosmos 3 Edge (4B) used for on-device vision reasoning and robot policy deployment.

Skeptics say today’s self-improvement systems still work best inside narrow, engineered harnesses rather than through open-ended discovery. TechTalks argued that reinforcement learning with verifiable rewards is strongest when answers are known in advance and feedback arrives only at the end, a sparse setup that can reward outputs without truly evaluating reasoning; that concern is echoed by arXiv:2609.24972, “RRSI: Regularized Recursive Self-Improvement of Agent Harnesses,” which explicitly “addresses overfitting in agent harness optimization by regularizing both the proposal and selection of edits to ensure improvements generalize to unseen tasks and benchmarks.” Even work on test-time gains remains bounded by engineered procedures: arXiv:2609.35748, “Improving Test-Time Scaling with Adaptive Looped Transformers” introduces “TaH2,” which “selectively applies additional computational iterations to tokens that benefit most from them,” improving “how language models trade off accuracy and compute at test time” rather than demonstrating open-ended innovation.

The deeper objection is that these systems still fall short of robust general intelligence when they leave curated settings. A Stanford-led team including Fei-Fei Li reported on May 18, 2026 that in every frontier multimodal model tested, “when forced to actively move through a 3D environment rather than interpret a pre-composed image,” the systems chose wrong actions with high confidence; that fits earlier evidence that MMSI-Bench found OpenAI’s o3 at roughly 40 percent on multi-image spatial reasoning versus a 97 percent human baseline, even as Anthropic tracked autonomous agents with “100% coverage monitoring” and TaH2 improved “how language models trade off accuracy and compute at test time” through selective extra iterations.