Artificial intelligence is no longer just a tool built by humans. It is now a co-creator.
Anthropic recently dropped internal data revealing that more than 80% of the code merged into its core repository is authored by Claude. Think about that number. Before Claude Code launched in early 2025, that figure sat in the low single digits. Now, engineers aren't spending hours typing out routine logic. They are directing, reviewing, and letting the model handle the heavy lifting.
This shift points directly toward recursive self-improvement. That means an AI system capable of designing and developing its own successor without human intervention. We aren't fully there yet, but the trajectory is moving fast.
The Productivity Explosion Inside Anthropic
Lines of code alone can be a noisy metric. But the reality on the ground is undeniable.
During Anthropic's first few years from 2021 to 2024, the amount of code an average engineer merged per day stayed flat. Then 2025 hit. Claude started running code instead of just spitting out text suggestions for copy-pasting. By mid-2026, the typical engineer was merging roughly eight times as much code per day compared to 2024 benchmarks.
The bottleneck shifted from writing code to reviewing it. When an AI can autonomously execute tests, fix bugs, and refactor entire modules over hours instead of minutes, the speed of engineering changes completely.
Beyond Coding: The Rise of Autonomous Research Loops
Writing functional software is one thing. Doing open-ended research is entirely different. Or at least, it used to be.
In a recent internal experiment, Anthropic gave Claude-powered agents an open problem in AI safety. The goal was to figure out if a weaker model could reliably supervise a stronger one. The agents were left entirely alone to propose hypotheses, run experiments, test variables, and share findings across parallel loops.
The results caught researchers off guard. Human teams working on similar problems for a week typically recovered about 23% of a specific performance gap. Claude's autonomous agents closed 97% of that gap after running for roughly 800 cumulative hours and consuming about $18,000 in compute.
Humans still set the initial problem and defined the scoring rubric. But the actual execution? That belonged entirely to the machine.
The Bottleneck of Taste and Direction
If AI can write 80% of the code and run complex experimental loops, why hasn't recursive self-improvement already happened?
The answer comes down to intent. Claude is exceptionally good at chasing a target once a human points it in the right direction. Give it a clear specification, an optimization goal, or a set of correctness checks, and it will iterate until it wins.
However, picking what to work on next requires taste. Deciding which architectural leap matters, which safety trade-off is acceptable, or what product direction serves humanity best are choices that still require human judgment. The gap between executing a brilliant experiment and inventing the next paradigm of intelligence remains wide.
What This Means for the Rest of Us
We are entering a strange phase of technological development. The companies building AI are now using AI to build the next iteration. This creates a compounding loop of capability gains.
When a model like Claude Mythos Preview can achieve a 52x speedup on an optimization task that takes a human researcher up to eight hours to notch a 4x gain, the timeline for future capability jumps collapses.
Anthropic is publicly calling for caution while aggressively racing forward. They want the world to consider safety pauses even as their internal metrics scream that autonomous software development is working too well to ignore.
The next generation of AI won't be crafted in a cleanroom by tired developers typing on mechanical keyboards. It will be spun up by code that wrote itself.
This video provides a deep-dive breakdown into Anthropic's research on recursive self-improvement and how Claude handles internal coding workflows.
