YuMoney Hands AI Agents the Review of 411 Comments on a Single Documentation Fix
At YuMoney, one pull request (a change request for documentation) collected 411 reviewer comments. The team wrote an instruction for an AI agent that sorts them into three piles: the answer is already in the document, a check against the code is needed, or the question is genuinely open. The agent drafts replies, but an analyst publishes them and makes the decisions. The logic is that routine work with clear rules is worth handing to agents, while the choice of solution stays with a person.
AI-processed from Habr AI; edited by Hamidun News
Imagine you wrote a technical proposal, posted it for approval, and colleagues left 411 comments on it. That is what one pull request at YuMoney looked like. A pull request is a request to merge changes: the team discusses edits there, and under the Docs-as-Code approach (documentation lives in the repository next to the code), technical descriptions are approved the same way. Arina, a systems analyst at the company, told Habr how her team handles such backlogs: instead of chatting with a neural network, they built a set of AI agents with repeatable scenarios. The author herself calls the 411-comment case exceptional, but it clearly shows where automation pays off.
From prompt to process
The team's main advice: don't start with the prompt, start with the process. An analyst's work usually begins with a Jira task that lacks requirements. You need to look for context, talk to adjacent teams, and check whether all scenarios are covered. Some steps in this cycle require expertise, and others can be formalized. Architectural decisions are not delegated to the analyst. But checking a document against conventions, finding similar solutions in the project history, or drafting a description can well be automated.
YuMoney uses the corporate tool OpenCode and four basic concepts. An agent receives a task, goes through sources, and produces a result. A skill is an instruction for a specific task: the order of actions and sources are described once, after which the agent only needs to be told which document to check. A command is a quick launch of a ready-made scenario, for example "describe the changes in the commit." The Agents.md file holds the basic project context: how the system is built, available resources, and rules, so the agent doesn't have to reconstruct them on every request.
What exactly was delegated to agents
The first scenario is searching for historical context. A project may return to work years later, when the code has changed but the decisions survive only in old discussions. The skill gathers past tasks, discussions, and the current implementation in the code. The analyst gets not a finished solution but a starting point: what has already been done and what constraints exist.
The second scenario is AI in the role of a critic. The analyst first formulates their own hypothesis, and the agent stress-tests it. It asks what happens if an external system is unavailable, how a timeout behaves, which alternatives were not considered, and which edge cases were left out.
The third and fourth are documentation and review. The agent knows the document's structure and formatting rules, so it prepares a draft, looks for typos, duplicates, and logical gaps, and builds a diagram from the text or text from a diagram. For review, the team made a skill that sorts comments. The agent divides those 411 into three groups: the answer is already in the document, a check against the code is needed, or the question is genuinely open and could change the decision. It writes draft replies, but the analyst publishes them and makes the decisions.
The fifth scenario is updating documentation after a release. The agent compares the technical solution with the code and tasks and produces a list of discrepancies. The analyst confirms which edits are needed, and only then does the agent update the document.
Rules so automation doesn't become one more chore
The team formulated three principles. First: one skill, one task. If a scenario has grown and performs several operations without intermediate checks, it should be split. Second: record intermediate results in the project's external memory so you don't have to reconstruct context from an endless chat. Third: keep skills, commands, and instructions in a shared repository organized by area. At first everyone built tools for themselves, but shared scenarios for recurring tasks that passed review proved more effective.
What stays outside the frame
The author notes a risk: if you constantly rely on ready-made answers, you can stop understanding the details. Hence the rule: your own hypothesis goes first, the agent's check goes second. She sees the analyst's value in responsibility for the result, the quality of the input data, and the ability to evaluate the AI's work.
At the same time, the text contains no savings figures: no hours saved, no token usage, no infrastructure cost. The only concrete number is 411 comments. Some skills, such as searching historical context, are still in development or testing. The material is tied to the announcement of the YuMoneyDay conference, so this is a report on experience, not a measured benchmark.
The practical takeaway is simple. Agents are worth deploying on tasks with clear rules and little context: sorting comments, checking documents against code, finding duplicates. Leave the decision and the publication to a human. Set up a shared skills repository so that successful scenarios don't remain personal notes.
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
What are the key details on From prompt to process?
The team's main advice: don't start with the prompt, start with the process. An analyst's work usually begins with a Jira task that lacks requirements. You need to look for context, talk to adjacent teams, and check whether all scenarios are covered. Some steps in this cycle require expertise, and others can be formalized. Architectural decisions are not delegated to the analyst.
What exactly was delegated to agents?
The first scenario is searching for historical context. A project may return to work years later, when the code has changed but the decisions survive only in old discussions. The skill gathers past tasks, discussions, and the current implementation in the code. The analyst gets not a finished solution but a starting point: what has already been done and what constraints exist. The second scenario is AI in the role of a critic. The analyst first formulates their own hypothesis, and the agent stress-tests it.
What are the key details on Rules so automation doesn't become one more chore?
The team formulated three principles. First: one skill, one task. If a scenario has grown and performs several operations without intermediate checks, it should be split. Second: record intermediate results in the project's external memory so you don't have to reconstruct context from an endless chat. Third: keep skills, commands, and instructions in a shared repository organized by area. At first everyone built tools for themselves, but shared scenarios for recurring tasks that passed review proved more effective.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.