Anthropic’s latest research offers some of the clearest evidence yet of how far this transition has progressed. The company describes the concept as recursive self-improvement: a process in which AI systems contribute to building better versions of themselves.
Fully autonomous self-improving systems remain some distance away. Yet many of the foundations required for such systems are already visible inside real-world AI development pipelines.
AI enters its own development loop
For decades, AI development followed a relatively straightforward structure. Humans defined problems, wrote code, trained models and evaluated outcomes. Every iteration depended on human effort, limiting progress to the pace at which teams could move through the cycle.
Anthropic’s findings suggest that structure is beginning to change.
The company reports that more than 80 per cent of the code it now merges into its codebase is written by its own AI models. This is not a marginal shift. It means AI is already responsible for the majority of the technical work involved in building AI systems.
The implications extend beyond productivity.
When AI-generated code becomes part of the systems used to train, improve or deploy future models, the technology enters a feedback loop. AI is no longer simply being developed. It is helping shape the conditions that influence its own future development.
The role of engineers is changing alongside it. Rather than writing code line by line, developers increasingly define objectives, review outputs and guide system behaviour. The focus shifts from execution to oversight.
Speed, scale and compounding gains
The most immediate impact of this transition is acceleration.
Anthropic says output per engineer has increased roughly eightfold since 2024. Internal estimates suggest productivity gains of around four times across teams. These figures are significant not simply because they reflect greater efficiency, but because they create compounding effects.
As AI takes on more of the development process, it speeds up the work required to build better AI systems.
Faster code generation enables faster experimentation. Faster experimentation leads to quicker model improvements. Better models, in turn, accelerate development even further.
The result is a feedback loop where each iteration helps improve the next.
External research from Model Evaluation and Threat Research (METR) points to a similar trend. According to its findings, the complexity of tasks that frontier AI models can handle has been doubling roughly every seven months.
Anthropic’s own data suggests these systems are also becoming capable of operating independently for longer periods. Tasks that once took only minutes now extend to hours and, increasingly, full work sessions.
From coding assistant to experimental researcher
The next stage of this evolution goes beyond code generation.
Anthropic describes emerging systems capable of proposing hypotheses, conducting experiments, evaluating outcomes and iterating towards better solutions.
This is important because it mirrors the basic mechanism required for self-improvement. Any system capable of improving itself must be able to generate changes, test them and determine whether those changes actually represent progress.
In one example cited by Anthropic, AI agents working on a research problem recovered nearly the entire performance gap between a baseline and an optimal solution, outperforming human researchers operating under tighter time constraints.
The system was not simply following instructions. It was navigating a defined problem space with a degree of independence.
Yet the limits remain equally important.
Humans still define the objectives, establish evaluation criteria and determine which problems are worth solving. AI operates within a framework created by people.
In other words, current systems are demonstrating the mechanics of self-improvement, but only within boundaries established by humans.
They can optimise within a problem. They cannot yet determine what the problem itself should be.
Microsoft’s approach: Controlled self-improvement
While Anthropic focuses on how self-improving behaviour is emerging naturally inside AI development, Microsoft is attempting to formalise the process.
Its Frontier Tuning initiative aims to create systems that continuously improve through real-world usage while remaining within tightly controlled environments.
At the centre of the approach are reinforcement learning environments where AI models learn from an organisation’s own workflows, data and evaluation signals.
These environments function as controlled spaces where models can improve over time without directly affecting production systems.
The principle is straightforward: learning should not stop once a model is deployed. Instead, systems continue adapting based on how they are used, becoming increasingly specialised to specific organisational needs.
What differentiates Microsoft’s approach is its emphasis on governance.
The learning loop exists, but it operates within compliance frameworks, access controls and predefined evaluation systems. This makes the process predictable, measurable and auditable.
Microsoft is not building fully autonomous self-improving AI. It is building constrained self-improving systems where the feedback loop exists but remains carefully managed.
Google focuses on the missing pieces
While Anthropic and Microsoft are demonstrating what is already possible, Google DeepMind is concentrating on the problems that still need to be solved.
In an interview with Axios, DeepMind chief executive Demis Hassabis identified two major barriers to fully autonomous self-improving systems
The first is the absence of robust world models. For a system to improve itself effectively, it must understand how its actions translate into outcomes. In environments such as chess, this is relatively simple because the rules are fixed and consequences are clear. Real-world environments are far more complex. Outcomes are uncertain, conditions change constantly and cause-and-effect relationships are often difficult to model.
The second challenge is verification. Even if a system generates a potentially better solution, it still needs a reliable way to determine whether that solution actually represents an improvement.
In coding, tests can provide answers. In mathematics, proofs can serve as verification. In many real-world situations, however, there is no immediate or objective signal that defines success. These limitations help explain why self-improving behaviour is emerging unevenly.
Progress is advancing rapidly in areas such as software development and research workflows, where outcomes can be measured precisely. It remains far more limited in environments where success is harder to define.
The rise of organisational learning loops
Even within these constraints, self-improving systems are already beginning to affect businesses.
As AI becomes embedded in workflows, it generates outputs that can be evaluated, refined and fed back into future iterations.
This creates feedback loops not only for AI models but also for organisations themselves.
Over time, companies can build systems that continuously improve based on their own operational data and internal processes.
The competitive implications are significant.
Historically, companies gained advantages through better tools, larger workforces or greater financial resources. Increasingly, advantage may come from how effectively organisations create and manage their own AI learning loops.
Two businesses using the same underlying model can achieve very different outcomes. One may treat AI as a static tool. Another may continuously refine it using internal data, feedback mechanisms and evaluation systems. Over time, the gap between those approaches could widen substantially.
This also changes the nature of software.
Rather than existing as fixed systems updated periodically, AI-driven systems become dynamic and adaptive. Their performance depends not only on how they were designed but also on how effectively they learn from usage.
What comes next?
The developments described by Anthropic, Microsoft and Google point towards a common direction. Self-improving AI is not arriving as a single, dramatic event. Instead, it is emerging gradually through a series of feedback loops embedded inside development pipelines, enterprise systems and research environments.
For now, those loops remain constrained.
Humans still define objectives, establish boundaries and determine what success looks like. AI systems can optimise within those frameworks, but they cannot yet create entirely new ones.
That distinction remains crucial.
Yet the underlying trend is becoming harder to ignore.
AI is increasingly participating in the processes used to build, refine and improve future AI systems. The feedback loops are already forming. The question is no longer whether self-improving AI is possible. It is how far these early forms of self-improvement can evolve before the remaining technical barriers begin to fall.
