Yandex B2B Tech has opened access to autonomous AI agents in the SourceCraft development platform: they operate under their own account as independent team members — they independently prepare the working environment, write and check code, run tests, and return the result for review, while reaching out to people when data is lacking or approval is needed. According to a McKinsey estimate cited by Yandex, such agents reduce the time spent on development tasks by an average of 11.2%, and the volume of rework by 6.8%.

image
image

What happened

The agent is assigned as the executor of a task: the easiest way to do this is directly in GitLab discussions, but tasks can also be accepted through the SourceCraft web interface, VS Code editor, command line, and messengers. Upon receiving a task, the agent independently prepares the working environment, makes changes to the code, runs tests, and sends a link to the changes to the original task; if the reviewer leaves comments, the agent refines them. The platform allows connecting your own agents, including those built in Yandex AI Studio.

Context

The launch continues the shift from AI autocomplete to autonomous executors integrated into the existing development process. Yandex Cloud CTO Ivan Puzirevsky formulates the principle as follows: infrastructure should seamlessly scale existing processes, rather than rebuilding them from scratch for each new agent — therefore, the agent is integrated into the familiar GitLab cycle, not a separate chatbot. The bet is also based on Yandex's internal experience: according to the company, 73% of its developers regularly use AI, more than half of the new code is created by generative models, and the 75/75/75 program requires raising both indicators to at least 75%.

Why this matters for the industry

For the industry, the launch is a signal that the agentic product is maturing: the "agent as executor" pattern is packaged into an existing workflow, not a separate chatbot, and the entry point is in the task tracker, not in the IDE. Technically, what matters is not the model, but orchestration: multiple task-setting channels and support for third-party agents from Yandex AI Studio indicate a bet on interoperability, not a single closed model. The planned unified subscription for SourceCraft, Alisa AI Pro, and Yandex AI Studio models adds a market layer — centralized accounting and limiting AI costs turn agents into a measurable line item in company expenses. If the mechanics are confirmed in practice, the "agent-executor in the task tracker" pattern will be quickly copied by other platforms, and competition will shift from the models themselves to the quality of orchestration, access rights, auditing of agent actions, and measurable quality metrics.

Why this matters for users

If your team works in SourceCraft or GitLab, you can try the agent without changing tools: assign it as the executor of a bug in a task discussion — it will prepare the environment, send a fix for review, and you will either accept the result or leave comments for refinement. You can communicate with the agent from familiar places — the web interface, VS Code, command line, and messengers, and a specialized agent for your tasks can be built in Yandex AI Studio. The barrier to experimentation is minimal, but the effect should be measured by your own metrics: before scaling, it makes sense to measure the baseline — cycle time, the share of rework, and review load — and compare it with the indicators after transferring routine tasks to agents.

What is still unknown / limitations

The announcement does not include benchmarks, an indication of which model is behind the agents, or data on price, latency, and the share of tasks that agents close without escalation to people. The claimed effects are attributed to a McKinsey report cited by Yandex itself: these are average values without disclosed methodology, task distribution, and baseline, so transferring them to a specific team should be treated as a hypothesis, not a measurement. Yandex's internal metrics — the share of developers with regular AI use and the share of code created by generative models — are not verifiable from the outside, and the methodology for counting such code is not disclosed and may mix autocomplete with the work of autonomous agents. There are no independent observations yet, for example on the share of reviews accepted without comments, as a result of the agents' work.

Sources

Author

Look at AI, editorial team