Anthropic says recursive self-improvement is not a closed loop yet, but its latest Institute essay makes the question harder to dismiss. Claude now writes more than 80% of the code merged into Anthropic’s own codebase, and the company says model autonomy on software tasks has stretched from minutes to half-day work. That does not prove AI can design its successors alone. It shows why frontier labs are treating AI-assisted AI development as an engineering problem rather than a thought experiment.
Table of Contents
The short version
- Anthropic says more than 80% of the code merged into its codebase in May 2026 was authored by Claude.
- The company argues that AI systems are taking over more of the implementation and experiment-running work inside AI development.
- Anthropic cites a rough trend where models double the length of tasks they can complete about every four months, including Claude Opus 4.6 handling 12-hour software tasks.
- The useful caveat is that choosing research directions, judging failures, and setting safety boundaries remain much harder than generating or optimizing code.
- Hacker News readers focused less on the benchmark numbers and more on incentives, regulation, open source, and whether Anthropic is describing safety policy or market power.
What happened
Anthropic’s Institute published “When AI builds itself,” an essay about recursive self-improvement and the practical work that might lead toward it. The company does not claim that Claude can autonomously create a better Claude. Its narrower claim is that AI systems are now doing a growing share of the work that improves future AI systems.
The concrete figures are still striking. Anthropic says that, as of May 2026, more than 80% of code merged into its internal codebase was authored by Claude. Before Claude Code’s research preview in February 2025, the share was in the low single digits. The company also says lines of code merged per engineer per day rose roughly eightfold from 2024 to the second quarter of 2026, while warning that code volume can overstate real productivity.
That last warning matters. More code is not automatically better code, and research prototypes do not carry the same maintenance burden as production systems. The signal is not that engineers became eight times more valuable overnight. The signal is that AI labs are reorganizing some daily work around agents that can write, run, and revise code inside the research loop.
Why recursive self-improvement is worth watching
Recursive self-improvement is worth watching because the early version may look mundane: agents writing code, running experiments, fixing bugs, and handing results back to researchers. A full self-improving AI would need to choose its own research agenda and reliably produce better successor systems. Anthropic’s evidence stops short of that, but it points to a path where the implementation layer becomes increasingly automated.
Anthropic gives several examples. Claude Opus 3, released in March 2024, could complete software tasks that took humans about four minutes. A year later, Claude Sonnet 3.7 handled tasks around an hour and a half. Anthropic says Claude Opus 4.6 managed 12-hour tasks, and METR found Claude Mythos Preview could work for at least 16 hours on long-duration task benchmarks.
For builders, the shift is practical. Once an agent can handle multi-hour tasks, teams need better review queues, evaluation harnesses, permissions, rollback paths, and cost controls. The question stops being “can it write a function?” and becomes “which work can run overnight without quietly making a mess?” For more coverage of this kind of AI tooling shift, the Diligesker IT & AI archive tracks related developer and model infrastructure stories.
What does recursive self-improvement change for builders?
Recursive self-improvement changes the builder problem from prompt quality to workflow design. If AI agents contribute to the systems that train, evaluate, and deploy later AI models, teams need to treat agent output as part of the development supply chain. That means reproducible experiments, audit trails, human approval points, and clear rules for when an agent is allowed to modify infrastructure or model code.
Anthropic’s essay separates execution from judgment. Claude appears strong at well-scoped implementation and optimization work. The company says Claude Opus 4 achieved about a 3x speedup on a small model training optimization task in May 2025, while Claude Mythos Preview reached about 52x by April 2026. A skilled human researcher, according to Anthropic, would need four to eight hours to reach 4x on that task.
Those numbers are impressive, but they do not remove the hard part of research management. Someone still has to decide which experiments matter, which shortcuts are acceptable, and which apparent gains are artifacts. Teams that adopt coding agents without stronger evaluation will get more output before they get more understanding.
What Hacker News readers are arguing about
The Hacker News discussion around Anthropic’s essay is skeptical, political, and occasionally useful. The thread has hundreds of comments, and the loudest split is about trust. Some readers treat the essay as a serious warning from a frontier lab that sees capability gains from the inside. Others read the same safety language as a pitch for regulation that could protect incumbents and hurt open source or smaller competitors.
One practical line of debate is about incentives. Several commenters asked whether a company that benefits from being near the frontier should be trusted to define the rules for slowing the frontier. Others argued that a credible pause would require international coordination because one lab slowing down alone would not solve the race dynamic. That maps closely to Anthropic’s own policy section, which says a pause would need multiple well-resourced labs, clear triggers, verification, and an adjudication process.
The developer-focused comments were more grounded. Some readers described agent workflows where one model generates experiments and another analyzes test harness feedback. Others pushed back on lines of code as a weak proxy for productivity. The useful takeaway is that the community is not arguing only about whether Claude can code. It is arguing about who gets to control the loop when coding agents start improving the tools used to build future models.
The practical read
Recursive self-improvement is not something most teams can verify from a headline. The safer read is to watch the intermediate indicators. Can agents complete longer tasks without losing context? Can they improve experimental code without gaming the benchmark? Can teams audit the changes afterward? Can labs slow or pause high-risk work in a way that competitors, regulators, and the public can inspect?
For developers and AI product teams, the near-term action is boring but necessary: build agent workflows as if they will become part of the critical path. Keep human review on high-impact changes. Measure accepted code against tests, incidents, and maintenance load instead of code volume alone. Treat AI-generated experiments as useful proposals until independent checks say otherwise.
For policymakers, Anthropic’s essay is a reminder that safety debates are moving from abstract capability forecasts into operating details. Verification is the hard part. Training runs are easier to hide than missile silos, and compute, data, and software tooling have legitimate uses outside frontier model development. A pause that cannot be verified will not slow the actors that matter most.