Agentic Engineering
Recursive self-improvement: are we there yet?
Not yet, says Anthropic. One of its co-founders thinks 2027 is the most likely year, and its own count has none of its AI research fully autonomous as of August. Our way of working changes over time too, and a person decides each change.
In plain English
You may have heard of "self-improving AI". Anthropic uses the term recursive self-improvement for an AI system that designs and builds its own successor without human help (opens in a new tab). Anthropic says that has not happened yet.
Our process is not that. Agents help us find and test ideas. They do not decide which ones we keep. A person approves each change.
The page address says self-improving. We mean the loop below, which a person runs.
How an idea moves
- An idea comes up. A better rule, a new check, a change to how we review or build.
- We check it against data. Does the problem happen? How often? What did it cost? Is something already in place?
- We argue the other side. We write down the strongest case against the idea, and what data would change the verdict.
- A person decides. Build it, add a small rule or note, say it is already covered, or drop it.
- We build it or drop it. A change we keep is reviewed before it goes live.
- We record the result. The change and the reason go into a written change log.
What the loop changes
It changes how we work: our rules for agents, our review prompts, our checks and our scripts. A change that reaches a product still goes through the review and tests described on how we work and peer review.
What this page does not show
We do not publish counts of ideas reviewed, built or dropped. This page describes the steps, not a score.
Approval is a separate choice
Two questions often get mixed up. Can a system improve itself? Does a person approve each change? The answers are independent. A system could improve itself with or without a person approving.
We keep a person approving each change, on purpose. That is a choice about control.
What Anthropic says about people in the loop
The essay linked above describes how Anthropic now works with its own AI. It is Anthropic's account of its own work, not a study of teams like ours. Three points explain why we keep a person in charge.
- AI review catches real bugs. An automated Claude reviewer reads proposed code changes before they merge. In a look back, Anthropic estimates it would have caught roughly a third of the bugs behind past incidents on claude.ai. That is why AI review is part of our process. It does not replace a person's decision.
- The bottleneck moves. Anthropic says that speeding up one part of the work often shifts the bottleneck to the parts that did not speed up. It names human code review as a new bottleneck, and it says that spotting and fixing bottlenecks may become the most important skill for any organization. Our loop is how we look for ours.
- People still choose. Anthropic says that, for now, people have the edge in judgment: which problems matter, which results to trust and when an approach is a dead end. In our loop, that is the step where a person decides.
The second source is news. On 2 November 2026, TechCrunch reported on Anthropic's revised filing for a public stock offering (opens in a new tab). The filing says Anthropic has in the past not noticed a significant speed-up in its own development until after it happened. It also says Anthropic is considering changes to its safety policy to keep humans in control of frontier model development. TechCrunch notes that skeptics might point out that such a filing is both disclosure and marketing, so we read it as one company's statement, not as an outside check.
Both sources point the same way as our loop: look at the data before you give a process more work, and keep a person deciding.
How fast this is moving
These figures are Anthropic's, reported by the press. We pass them on as context, not as a measure of the field.
Anthropic published a chart of how much of its own AI research and development is done by AI. The Wall Street Journal reported on it on 30 October 2026 (opens in a new tab), behind a paywall. At the level where AI does most of the work and a person supervises, the chart shows 1% in March 2026 and 26% in August 2026. A year earlier, AI led none of that work. As of August, none of it was fully autonomous. This is one company's count of its own work.
The same article reports a forecast from Jack Clark, an Anthropic co-founder. In May he thought an AI that builds its own successor was likely by the end of 2028. He now thinks 2027 is the most likely year. That is one person's view. Anthropic's essay, linked above, gives no date and says this is not inevitable. However the numbers move, a person approves each change here.
Two ideas the loop handled
These two ideas come from our own work, as of November 2026. Each rests on a written record. The records live in our private repositories, so we cannot link them here.
Kept: repeat the self-check before each later review round
Before a change goes to review, we check our own work by hand. The idea was to do that again before every later review round. The data: in a long plan review, later rounds kept finding sentences that earlier fixes had added and that claimed more than their evidence showed. Our rule covered only the first round.
The case against: one round in that review had to check a real bug fix, so the saving was smaller than the number of late rounds suggests. A person said yes anyway, as a small rule, not a new tool. It passed review with only wording notes and went into our working rules on 5 November 2026.
Dropped: check the second reviewer's credit before every review
We use a second AI reviewer from another provider. The idea was to check that reviewer's credit before each review starts, so a review is never quietly weaker when that reviewer is out. Our notes from early November 2026 say the provider documents no way to read the balance, a failed second reviewer costs about a second and no review round, and the first round already shows when that reviewer is out.
A person said no: when the second reviewer is out, the review goes on with the first one alone. The idea came back twice more, and the written answer stood. It would come back if the provider offered a balance check, or if a person wanted reviews to pause when the second reviewer is out.
Frequently asked questions
What is recursive self-improvement?
An AI system that improves its own design, so the next version improves itself again. Anthropic adds a condition: the system does it without human help.
Does using AI to improve our own work count as self-improvement?
Not in Anthropic's sense. An AI that suggests a better review prompt, which a person then accepts or drops, does not build its own successor without human help.
Why keep a person approving each change?
An agent can propose a change, test it and argue against it. It cannot answer for the result. A named person can, so we accept the slower pace.
Related reading
How we know the AI is good enough, our process, and agentic engineering, not vibe coding.