AI is no longer just a tool that helps humans build software-it is increasingly the main developer in the loop. According to a new report from Anthropic, the company behind the Claude family of models, artificial intelligence is already designing and improving the very systems that power it. And in this new landscape, the limiting factor may not be the models themselves, but the humans supervising them.
In its study titled “When AI Builds Itself,” Anthropic describes a development pipeline in which Claude has become a central engineering collaborator rather than a passive assistant. The model drafts code, runs experiments, interprets results, and even helps plan research directions. People are still in charge of deciding what to build and whether the outputs are safe and correct-but much of the day‑to‑day work in between is now handled by AI.
One of the most striking data points from the report: Claude now generates more than 80% of the code that ends up merged into Anthropic’s own codebase. Just a year earlier, before the release of the specialized Claude Code tools in early 2025, that figure was in the “low single digits.” In other words, in a very short time AI has gone from occasional helper to dominant code author inside a cutting-edge AI lab.
Anthropic claims this shift has had a dramatic impact on productivity. Since 2024, engineers at the company have seen their code output increase by roughly a factor of eight. That doesn’t mean eight times as many people or eight times as many hours worked; instead, much of that extra throughput comes from delegating routine and semi-routine programming tasks to Claude. Humans define the goals, set constraints, and review the results. The model fills in most of the implementation.
Beyond just writing code, Anthropic says Claude is increasingly involved in the full research cycle. The model is used to propose experiments, generate configuration files and scripts, run or orchestrate those experiments, and help analyze the resulting data. It can suggest follow‑up tests, new parameters to explore, or alternative architectures-acting in a role that, in traditional labs, would have required a team of research engineers and analysts.
This dynamic points toward what researchers call “recursive self‑improvement”: AI systems that help create better versions of themselves. Anthropic stops short of claiming that full self‑improving loops already exist, but the report argues that current practice is an early, practical step in that direction. An AI that designs training runs, writes the code for new features, and speeds up iteration is, in effect, participating in the design of its own successors.
Paradoxically, Anthropic argues that humans are now starting to look like the bottleneck in this process. The models can produce code, documentation, and experiment configurations at a pace that far outstrips human review. The slowest stages become specification (deciding what to build) and evaluation (checking that it works and is safe), both of which still rely heavily on human judgment.
This creates a new kind of speed limit for AI development. It’s not primarily about GPU shortages or algorithmic breakthroughs; it’s about how quickly humans can reliably say “yes” or “no” to the output of an endlessly energetic machine collaborator. As Claude and similar systems get more capable, the gap between what they can propose and what humans can thoroughly vet is only likely to widen.
In practice, Anthropic describes workflows where an engineer sets high‑level objectives-such as improving a component, adding a feature, or exploring a research idea-and then asks Claude to break that down into subtasks. The model may then generate code for each subtask, propose tests, and even write documentation. The engineer’s role shifts from crafting every line to curating, editing, and deciding which directions are worth pursuing.
This shift changes the skill profile of AI engineering. Rather than being rewarded for brute‑force coding, developers gain leverage through problem formulation, critical thinking, and system-level design. The most valuable engineers are those who can ask precise questions, foresee failure modes, and quickly spot subtle issues in AI‑generated work. Technical “taste” and judgment become more important as the models absorb more of the mechanical labor.
At the same time, Anthropic’s findings raise obvious safety and governance questions. If AI systems are deeply involved in building future AI, how do we ensure that human values and constraints remain firmly in control? The more responsibility we delegate, the higher the risk that bugs, biases, or unintended capabilities propagate and compound across generations of models.
The report implicitly argues for keeping humans in the loop at the stages that matter most: choosing research agendas, setting safety policies, and approving changes to core systems. While Claude can suggest new architectures or training setups, people must still decide which of those ideas are ethically acceptable, economically sensible, and strategically aligned with long‑term goals. Delegation is powerful-but only if the supervising layer stays robust and independent.
Another emerging challenge is verification. When a model authors 80% of the code, traditional review methods can start to strain. It is difficult for human teams to exhaustively inspect every line in large, fast-moving codebases, particularly if they are under commercial pressure to ship improvements quickly. That reality may push companies to develop meta‑tools: AI systems that check, test, and formally verify the work produced by other AI systems.
This “AI supervising AI” pattern could further accelerate development, but it also risks creating opaque chains of reasoning where very few humans truly understand the full system. To counter that, organizations may need new standards for interpretability, documentation, and traceability of AI‑generated artifacts-so that future engineers can audit decisions made by both humans and machines.
From a broader economic perspective, the Anthropic study suggests that the cost structure of AI research and software development is already shifting. If one engineer working with a model like Claude can produce as much code as several engineers could previously, the marginal value of additional human coders may decline in some contexts, while the value of compute, data quality, and high‑level strategic thinking rises.
For developers and knowledge workers more generally, this creates both risk and opportunity. Routine coding, boilerplate generation, basic data analysis, and simple experimentation are rapidly being automated. At the same time, individuals who can orchestrate complex AI-driven workflows, design robust systems, and navigate ethical and regulatory constraints are likely to become more sought after, not less.
Anthropic’s framing also undermines the idea that AI progress is purely a story of smarter models. Much of the acceleration comes from better integration of existing capabilities into the research and engineering pipeline. Tools like Claude Code formalize the model’s role as a pair‑programmer, experiment planner, and research assistant. As similar tooling spreads, even organizations that do not build frontier models may see comparable gains.
Looking ahead, if the trend lines in the report continue, we can expect more parts of the AI lifecycle to be automated: data cleaning and labeling, architecture search, hyperparameter tuning, evaluation design, and even policy‑drafting around safety. Each step that becomes machine‑assisted further reduces the human labor required per unit of progress-intensifying both the pace of innovation and the stakes of getting oversight right.
The central tension highlighted by Anthropic is simple but profound: AI is making it easier and faster to build better AI, yet humans remain the gatekeepers of what “better” should mean. The models can propose countless paths forward; people must decide which ones to walk, where to slow down, and where to stop entirely. How effectively we manage that tension may determine not just the trajectory of AI research, but the broader impact of these systems on economies, labor markets, and society as a whole.