In an eight-day experiment reported in September, an AI research system discovered seven successive improvements to an agent’s own research software. The resulting gains carried over to tasks outside the selection process. A piece of the machinery used to do research had itself become a research subject. AIDE² paper .
This puts AI to work on the machinery of AI research itself. When improvements make a system better at finding the next improvement, the development process can begin to compound.
In AI is moving faster than our predictions , we argued that businesses may have less time to adapt than their forecasts suggest. Recursive self-improvement adds a reason to take that possibility seriously: AI is increasingly being applied to the work that produces AI.
The public evidence now includes limited recursive loops. It still leaves a substantial gap between those experiments and a system that independently builds ever more capable successors. Understanding that gap helps us judge which announcements change the outlook.
What makes improvement recursive?
A developer asking AI to fix a bug is ordinary AI assistance. An agent making demonstrably useful changes to its own tools is self-improvement. The process becomes recursive when the improved system participates in producing further improvements.
That can happen at different levels. An agent can change its prompts, memory or research software. A training process can update the model itself. The stronger scenario is a system that can manage the development of its successor, which then takes over the next cycle.
These distinctions matter. Better software around an existing model can deliver useful gains without making the underlying model more capable. More experiments can produce better results without proving that the system has become better at choosing worthwhile experiments.
The question to ask is what changed, what was measured, and whether the improvement survived the next round.
Five pieces of evidence worth watching
September: AIDE² improved the software around a research agent
The AIDE² preprint reports changes to the agent’s search strategy and management of its working context, with gains across four held-out benchmarks, including weather forecasting. This is improvement to the agent’s supporting software, often called its harness.
The authors also tested whether a self-improved agent was better at driving another improvement run. That comparison was inconclusive: there were only three runs per condition, and the final average scores were similar. The paper therefore supports transferable agent improvements, while leaving acceleration of the improvement process unproven. Results and limitations .
October: MASS carried improvements into model training
An 8 October preprint from Sakana AI and UC Berkeley goes a step further. In MASS, copies of the same model propose team workflows, execute them and judge the results. Selected examples then train an updated model, which takes over those roles in the next round.
After two rounds, the researchers report 1.2–1.6 times the base model’s benchmark score per output token across four research benchmarks. Separate external model judges assessed research outputs without supplying feedback to the training loop.
This is a limited experiment with model training inside the loop. It is also uneven: score per output token fell slightly on two software-engineering benchmarks. Two rounds, a small synthetic training suite and model-based evaluation cannot establish unlimited improvement. Paper , methods and results .
September: Anthropic measured AI taking a larger research role
Anthropic’s September report put Claude at the level of leading 26% of its AI research and development work as of August, up from under 1% in February. Leading means completing most of a task from a high-level instruction while a human supervises.
The same report says no measured subset was fully autonomous. Its index is an internal assessment, and Anthropic acknowledges limitations in using its own models to evaluate its systems.
This measures adoption inside a frontier lab. It tells us that the inputs to future AI development increasingly include work done by today’s AI. Anthropic’s measurement report .
September: OpenAI reported an automated research intern
On 6 September, OpenAI said its internal measurements showed it had reached its research-intern target: a system able to carry out well-defined research tasks under human direction, including work taking a skilled researcher several days.
Its stated goal of a fuller automated researcher by March 2028 remains a target. The reported capability is supervised research assistance, rather than a demonstration of autonomous successor development. OpenAI’s report .
April: agents designed and tested an AI-training method
Anthropic’s automated weak-to-strong research project asked whether a weaker model could help supervise a stronger one. Agents proposed and tested their own approaches within a problem and evaluation setup chosen by humans.
On the study’s task, they recovered 97% of a defined performance gap, compared with 23% for two researchers tuning selected existing methods over seven days. The agent run took five days, with 800 cumulative agent-hours across nine agents and about $18,000 in compute and API costs.
A transfer attempt using production training infrastructure produced an improvement within the noise floor. The experiment demonstrates useful autonomous research within a defined setting, while also showing how much the setting can matter. Research report .
The missing step is sustained acceleration
Together, these results show several parts of a possible feedback loop: AI doing research, changing its research tools, generating training examples and using updated systems in subsequent rounds.
They do not establish that each generation will become a better researcher quickly enough to make the whole cycle accelerate indefinitely. Lab reports also describe different things from controlled experiments, and neither should be treated as independent replication of the other.
The strongest next evidence would show repeated gains at comparable cost, on tasks excluded from development, with the improved system demonstrably better at finding further improvements. Independent reproduction would make that evidence more persuasive.
Compute, experiment duration and human judgment still matter. Suppose an experiment takes a week to run but only an hour to prepare. Making preparation ten times faster barely changes the elapsed time. Acceleration depends on removing the constraints that actually determine the pace.
What this changes for a business
A small company does not need its own AI research lab to feel the effect. If developers can produce useful upgrades faster, customers may face shorter intervals between capabilities becoming available and competitors putting them to work.
That makes long-lived assumptions expensive. A workflow rejected as technically impractical in one quarter may deserve another test in the next. Teams that preserve their examples, failure cases and success criteria can evaluate an upgrade much faster than teams starting again from a demonstration.
Our argument in Hyper-Agile applies here: faster implementation increases the importance of deciding what is worth doing and judging the result. A system that generates more proposals can also consume more review time. The useful measure is accepted work at an acceptable total cost.
Businesses can apply a modest version of this learning loop now. Let an agent propose changes to a workflow, compare the revised version with the current one on reserved examples, and keep improvements that reduce errors or effort. Preserve a rollback path and human approval for consequential actions. An agent should not be able to redefine success simply to make its latest version pass.
Our expectation is that AI-assisted AI development will keep shortening some capability cycles. The recent experiments make that expectation more concrete, while leaving the speed and ultimate limits open. The practical preparation is to build an organisation that can recognise a useful improvement, verify it and adopt it without restarting the learning process every time.