# The AI Self-Improvement Loop Is Closing. But Engineering Comes First

> Anthropic, OpenAI and Google DeepMind have each measured AI speeding up AI research, and all three find the automation sits in engineering, not research judgement. Enterprises running AI face the same boundary.

- URL: https://dperez.com/posts/ai-self-improvement-loop-engineering-first/
- Author: David Perez, Principal AI Strategist, AWS (https://www.linkedin.com/in/davidperezcaparros/)
- Published: 2026-11-07
- Tags: ai, agents, evals

Anthropic's amended IPO filing contains an odd pairing. In the risk factors, [the S-1 filed on 2 November](https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm) warns investors that the company _"may imminently meet"_ its own threshold for AI dramatically accelerating AI research. In the business section, it presents that same dynamic as _"one of the most significant structural advantages in frontier AI"_.

The dynamic is recursive self-improvement, or RSI. An AI system improves the process that produces its successor, so each generation arrives sooner, or better, than it would have without help. The idea is old. [I.J. Good sketched an "intelligence explosion" in 1965](https://linkinghub.elsevier.com/retrieve/pii/S0065245808604180). What changed this year is that Anthropic, OpenAI and Google DeepMind each published measurements of it.

Laid side by side, those measurements suggest RSI does not arrive all at once. It starts wherever "better" can be measured automatically, and today that means engineering. Deciding what to research, and judging whether a result matters, is still done by people. Whether that boundary moves is the question that shapes the next year of model development, and it has a direct counterpart inside every enterprise running AI.

## Three labs, one boundary

The labs disagree on how fast things are moving. Anthropic's essay, [_When AI builds itself_](https://www.anthropic.com/institute/recursive-self-improvement), says that as of May more than 80 percent of the code merged into its codebase was authored by Claude, while _"large performance gaps persist"_ when Claude has to exercise judgement in choosing goals. One route to the RSI threshold in its [Responsible Scaling Policy](https://www-cdn.anthropic.com/files/4zrzovbb/website/0bacdc8440ea96e62a8766d99ebe1d4eea6d5f3a.pdf) is a doubling in the rate of progress, _"as much progress in one year as one would see in two years at baseline"_, and the policy concedes the threshold _"has proven difficult to operationalize"_.

[OpenAI said on 6 September](https://openai.com/index/research-acceleration-view-inside-openai/) that it had built its promised automated research intern, a system for _"well-defined research tasks under human direction"_, and that its research organisation now runs 3.1 agent-workdays for every human workday. The same post says _"People still set our research priorities"_. Its next target is an automated AI researcher by March 2028.

Google DeepMind's [Gemini 4 Argon safety report](https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf) finds _"no significant end-to-end acceleration of AI R&D, AI-driven or otherwise"_. Most of its production code is AI-written, yet about 88 percent of pull requests still need repeated human steering. [Zvi Mowshowitz has argued](https://thezvi.substack.com/p/gemini-4-argon-the-model-and-safety) that the no-acceleration finding is hard to square with Google's own public talk of self-improvement.

All three are the labs grading their own work, and on acceleration they diverge. On where the automation sits they agree, and an independent [Princeton-led study](https://arxiv.org/abs/2607.27191) lands in the same place. Frontier agents given the research question behind two unpublished NeurIPS papers did the engineering unaided, and the original authors rejected both resulting papers.

The reason engineering goes first is mechanical. Models get better at a task fastest when training can score attempts automatically, and much of engineering comes with a scorer attached: the code compiles or it does not, the tests pass, the kernel runs faster, the loss goes down. Research judgement has no such signal. Nobody can check in an afternoon whether a research direction was the right one. Anthropic's essay puts the current state plainly: _"humans supply the goal, but they no longer need to supply the method"_.

## The boundary is what to watch next year

If the labs push that boundary even partly into research judgement, release cadence speeds up and the formal thresholds start to bite. Anthropic reports early movement. On hand-picked moments in real research projects, its best model has increasingly chosen a better next step than the human did. It calls this an early signal and warns it is not a like-for-like comparison. If the boundary holds, DeepMind's steady trend continues and gains keep coming mainly from compute.

Either way, the brakes are now real. After discovering in July that its agents had compromised its research infrastructure, OpenAI paused reinforcement learning on models meant for deployment for about two weeks. Anthropic's filing says it delayed pre-training a new frontier model by more than a month for security work, and that it is considering rules _"to require human review and approval of certain decisions"_ in research as models take on more of it. My expectation for the next twelve months is more frequent releases driven by engineering gains, more variance in their timing, and more of each lab's effort spent on controlling its own research environment.

## The enterprise version of the loop

Few companies will train a frontier model, but most will run something with the same shape. The model stays fixed. What improves is the system around it: prompts, retrieval settings, skill and rules files, tool definitions, test suites, the pipeline that scores outputs. The lesson from the labs carries over. An agent can improve that system only where someone has defined "better" in a form a machine can check.

The loop has five steps. Write an evaluation that scores the outcome you care about. Let an agent propose a change to the system. Run the change against the evaluation automatically. Keep it only if the score improves. Require a human to approve any change to the loop itself, including the evaluation, the agent's permissions and what counts as a pass. The first four steps are engineering, and teams with good evals can run them today. The fifth is the one Anthropic is now considering for its own research, for the same reason: a loop allowed to edit its own scoring will find the cheapest way to score well.

That makes evaluations the enabling infrastructure for enterprise AI over the next year. A team without them has nothing for an agent to improve against, and it will judge progress by activity. OpenAI's own appendix admits that easy indicators such as code volume are _"hard to interpret"_, and agent-hours are no better.

The same boundary applies inside the company, too. An agent can tune a retrieval pipeline against a well-built test set. Deciding which customer problem deserves a pipeline is still a human call, and an evaluation can only test for the goals someone wrote into it.

Two vendor-side moves follow. A provider's release schedule now depends on an internal loop that can pause on safety grounds at short notice, so make swapping models a configuration change and use your own evaluations to rerank candidates in a day. And ask vendors what they disclose about their own research acceleration, since OpenAI now argues that labs should be _"required to publicly track our progress toward RSI"_.

For leadership teams, the useful question for the next year is narrower than whether AI is about to build itself. Where in your AI estate can an agent already make something better, prove that it is better, and ship the change? And who signs off when the thing being improved is the loop itself?