Yandex automates Chromium updates with an LLM agent

Every four weeks, Yandex Browser moves to a new version of Chromium. End users never notice, but developers have to resolve thousands of code conflicts and fix just as many compilation errors. Yandex decided to automate this routine and built an LLM agent into its upstream synchronization process.

An update so large it eats up months

Each update cycle brings roughly 10 thousand commits from mainline Chromium and about 1,500 custom changes accumulated in Yandex's fork. The result: more than a thousand merge conflicts, several thousand compilation errors, and a mass of dependencies that stopped working with the new version. The process involves dozens of developers of varying skill levels. Total effort per cycle comes to several person-months. That is time the team cannot spend on features, optimization or user-reported bugs. It is pure routine that blocks development.

Why a simple LLM chat does not work

It is tempting to think you can feed an error into ChatGPT and get a fix back. In practice, that does not work. Yandex is not dealing with isolated errors; this is the regular update of a large fork with a deep history of changes. Thousands of problems have to be handled at once, taking into account the context of both upstream and the company's own modifications. A standard LLM query is useless here: the context is insufficient, solutions contradict each other, and every step needs validation. So Yandex went further. Rather than simply pasting errors into a chat, it built an LLM agent that understands the project's structure and can analyze and resolve conflicts autonomously.

How the agent resolves conflicts

The agent works in two stages. First, it analyzes the conflict: what changed upstream, what exists in the fork, why the two are incompatible, and which dependencies are broken. Second, it proposes a resolution that accounts for both sides. These are the main tasks the agent takes on:

• Resolving merge conflicts by analyzing changes from both sides
• Fixing compilation errors (type errors, linking errors, missing includes)
• Validating solutions: checking that nothing broke in neighboring modules

The agent is not perfect: complex cases still require manual work. But it handles 70-80% of the work on its own. Developers move into code-review mode instead of writing code, which is much faster and more efficient.

What this means for browser development

Vendors of large forks (browsers, operating systems, embedded kernels) now see a realistic path to less painful upstream synchronization. LLM agents are finding their place not in flashy demos and presentations but in routine industrial processes, where they actually cut costs. It is a sign of LLMs maturing in software development: not a replacement for the developer, but a smart assistant for routine tasks that frees up time for creative work and new features. Yandex is squeezing the maximum out of what has been added to open models: not hallucinations, but systematic work.