At a recent OFFZONE keynote, a figure from Yandex came up: around 40% agentic code. Whether that is a current number or a target was left unsaid.

Listen to enough conference talks and it becomes clear that the word “share” hides very different things:

• At the CIPR conference, T-Bank put the number at 35-40% of code currently written with AI.
• Raiffeisenbank Russia presented experiments where an agent wrote 80-99% of the code.
• Kaspersky reports that today 12% of pull requests are made with AI and 17% of suggestions are accepted by developers. The target is 30% of code within three years.

So one number describes the current state, another an intention, a third a specific metric, a fourth a set of isolated experiments. Nothing surprising here: the industry expects returns and is experimenting hard - with use cases, metrics, and levels of autonomy.

The share everyone quotes has a second half that almost never gets quoted alongside it: the review regime. What is closed mechanically, what is left to a human, and by what signal do you decide which pull requests nobody needs to read?

At Acclaim, the first review is handled by a model almost everywhere. To get there they mined three or four months of human-reviewed pull requests for about three hundred rules, condensed them into 66 categories, and split them across 12 subagents. A human still shows up in every case: nobody has settled how to tell a “complex pull request” from a simple one.

Another reason to hand part of the review to a model is the sheer scale of security checking. From Kaspersky’s CTO: “We have 275 million lines of code. I will never find enough SDL specialists to check my source code for vulnerabilities.” By his account the model continuously scans attack-surface code for vulnerabilities, and every week it finds something in the new code.

But machine review has a weak spot: when the agent writes both the code and the tests for it, the criterion stops being independent. Tests create the impression of a simple check - run them and you know. Yet Acclaim notes that “models love to fit the tests to the code,” so “everything comes out green.” Green tests turn out to be an ambiguous signal. Much like the empty log in the case I wrote about earlier: the block may never have been needed, or it may simply not have worked.

I ran into the same problem back at Samsung. In 2022 we used an LLM in the AI Developer Assistant project to generate unit tests for C/C++ code. For such a test you can mechanically verify the syntax, the compilation, and the call to the target function. But a test that checks nothing at all satisfies all three conditions just as well. Coverage tells you which code the test touched; mutation testing tells you whether it can catch a defect. Four years later, in 2026, Avito describes the same class of failure: the agent may report success when in fact “only the stubs passed the check.” Their defense is mutation testing.

At Yandex Bank the review regime is tied to release velocity: “we cannot afford to ship 20 releases a day until we have verified they are fine.”

And there is the opposite extreme. A high share does not always signal maturity - more often it signals the risk profile a company is willing to accept. As an acquaintance at a small company put it: shipping speed is everything, architecture and security come last, and “if a user files a bug, Claude will rewrite it.”

Taken on their own, these shares are generally not comparable. For each figure it helps to pin down three things:

• What went into the calculation.
• Whether it is an achieved number or a target.
• How the review regime works.

A sense of acceleration is not enough either. In an Avito pilot, a hundred engineers with access to frontier models reported speedups, but no statistically significant change showed up in objective development metrics.

So the target is worth setting as a pair: the share of agentic code and the regime by which it is accepted. Otherwise the percentage can grow on migrations, configs, and tests - work whose result is formally easier to verify - while the way of working stays exactly the same everywhere a mistake is expensive.