In the past few weeks, it has become almost impossible to keep up with all of the new features that the foundation model builders are shipping. The engineers are working so fast, by the time you configure the new new thing, there are a half-a-dozen new, new new things you have to deal with.
We’ve always lived on the exponential, but product releases are just about at “ludicrous speed,” and they are going to get faster yet. There’s a reason.
Anthropic built Cowork (its knowledge-worker productivity tool) using Claude Code. The entire product took a week and a half. Felix Rieseberg of Anthropic’s technical staff said they spent more time on product and architecture decisions than writing individual lines of code. Claude built Claude’s newest product. This fact should give you pause.
The Loop You Need to Understand
Recursive self-improvement has been debated in AI research circles for years. The science fiction version, where AI autonomously designs its successor in an exponential intelligence explosion, remains theoretical. If you’re interested in this “probable” future, the nice folks at ICLR 2026 have organized a workshop to study its algorithmic foundations. These conversations (and there are many of them each year) are always thought-provoking.
That said, a practical version of recursive self-improvement is already here, and it is already transforming the economics of innovation. It goes like this: AI dramatically accelerates the humans who build AI products, which produces better AI, which further accelerates the humans. This is the final “human in the loop” step before we get to the AI building itself. Even in its current form, it is very powerful.
The Tool That Builds Its Own Tools
A “skill,” as defined by Anthropic, is a modular instruction set that teaches Claude how to do a specific task or carry out an agentic workflow. Skills can generate a document format, review code for security flaws, research and write a proposal, triage insurance claims, etc.
In early March 2026 (last week), Anthropic launched a new tool called “skill-creator.” It’s a skill that builds other skills. Skill-creator has four modes: Create, Eval, Improve, and Benchmark. Four sub-agents run in parallel. One executes the skill against test prompts. One grades the output. One runs blind A/B comparisons between versions. One finds the failure patterns. The system splits test cases 60/40 into training and holdout sets, proposes fixes based on what broke, re-evaluates, and iterates up to five times automatically. Anthropic ran this against their own internal skills. Five out of six improved. The tool that builds skills now improves the skills it builds.
This is recursive self-improvement. And these skills are available now in Anthropic’s marketplace. Cisco published a software-security skill built on Project CodeGuard that scored 84 percent overall and nearly doubled Claude’s ability to write secure code across 23 rule categories. Anyone can install it. Anyone can fork it. Anyone can run evals against it and publish a better version tomorrow. The recursive loop has escaped the lab and is becoming an ecosystem.
Spotify Scaled It
In July 2025, Spotify integrated the Claude Agent SDK into their fleet management infrastructure. On their Q4 2025 earnings call, co-CEO Gustav Söderström told analysts that Spotify’s most senior engineers had not written a single line of code since December. They generate code and supervise it.
The system, called Honk, merges over 650 agent-generated pull requests into production every month. Spotify reports up to a 90 percent reduction in engineering time on complex code migrations. An engineer on their morning commute can open Slack on their phone, tell Claude to fix a bug, receive a new app build back in Slack, and merge it to production before arriving at the office.
Honk did not appear overnight. Spotify built fleet management infrastructure starting in 2022, standardized their developer platform through Backstage (Spotify’s open-source developer portal), and invested years in comprehensive test suites. The AI agent sits on top of world-class engineering infrastructure. The recursive advantage requires a foundation.
Research While You Sleep
On the AI training side, Andrej Karpathy’s autoresearch project strips the concept to its minimum viable form. One GPU. One training file. One metric: validation bits per byte. An AI agent modifies the code for a small language model, trains for five minutes, checks whether the loss improved, keeps or discards the change, and repeats. You go to sleep. The agent runs roughly 100 experiments overnight. You wake up to a log of everything it tried and a better model.
Karpathy’s framing is worth quoting: “… frontier AI research used to be done by meat computers in between eating, sleeping, having other fun, and synchronizing once in a while using sound wave interconnect in the ritual of ‘group meeting.’ That era is long gone.”
The repo has over 32,000 stars on GitHub. The community has already forked it for Mac and Windows.
The Pattern
Spotify’s Honk, Anthropic’s self-improving skills, Karpathy’s autoresearch. Three different domains, same machine. Human sets the objective. AI executes, tests, iterates. Human reviews and tightens the spec. Every cycle produces better output and better instructions for the next cycle. The compound effect is already showing up in earnings calls and shipping velocity.
These loops run 24 hours a day. They run while you sleep, while you commute, while you sit in meetings about whether to form an AI committee. Every organization running a recursive loop is compounding its advantage against every organization that is not. And the cycles are getting shorter.
The recursive advantage is not coming. It’s running.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.