Марія Бровінська AI Eng 10 November 2026, 08:55

Meet Alfred, Gandalf, PRODIGY and Cyrus. How AI agents are hiring people, writing code and automating work in Ukrainian IT

AI agents are gradually becoming part of the operational work of companies. They are delegated not only to search for information or prepare texts, but also to complete entire processes: processing customer requests, initial selection of candidates, access management, document preparation, reporting and marketing tasks. An agent can receive a request, check the necessary data in several systems, perform a number of actions and transfer the result to a person for approval.

These are no longer isolated experiments. According to a McKinsey study, 40% of surveyed companies with annual revenues of over $1 billion are scaling up the use of AI agents. Among the areas where they are used are IT, knowledge management, marketing, sales, and operational processes. At the same time, the spread of agents does not automatically mean a financial effect: only 37% of respondents reported at least some impact of AI on companies' EBIT.

Large companies are already delegating some of their customer interactions to agents. For example, Klarna’s AI assistant handled 80% of its support chats in 2025, a total of 31 million conversations. The company estimated its work to be equivalent to more than 700 full-time employees. Salesforce is also seeing a surge in agent usage: the company says that in the first half of 2025, the number of agents created and deployed among organizations in the study increased by 119%.

Ukrainian IT companies are moving in the same direction, although their cases often concern internal processes. dev.ua collected examples from Railsware, Headway, Sigma Software, and Ciklum and figured out which tasks can already be delegated to agents, how independently they work, and where human control remains mandatory.

Recruiting: the agent collects information, the person makes the decision

At Railsware, AI agents are used not only for engineering tasks. One of the most noticeable areas of their use is recruiting. Daryna Kuzmyk, the company’s Talent & Culture Lead, told dev.ua.

“In addition to development, the largest number of AI agents currently work in recruiting. The AI recruiter conducts initial pre-screening calls: collects incoming information from the candidate and makes a basic assessment of the suitability for the role. Agents on custom skills make the first assessment of the resume and structure the results of test tasks so that the recruiter immediately sees key competencies and experience,” she explains.

At the same time, the company does not delegate the final selection to algorithms. The agent can collect data, structure the resume, and prepare an initial assessment, but the decision regarding the candidate remains with the recruiter.

“We don’t have a process that works completely without a human. […] Agents analyze and structure, and the decision on each candidate is made by the recruiter. No candidate is screened out or hired solely by the algorithm,” Kuzmik emphasizes. According to her, the recruiter can save up to eight hours a week — effectively an entire working day. This time the team spends on deeper conversations with candidates, final interviews, and building long-term relationships with talent.

A similar approach is used at Ciklum. The company uses an AI model for initial resume analysis: it compares the candidate’s experience with the job requirements, assesses the profile’s suitability, highlights potential inaccuracies, and can suggest other roles that are a better fit for the person. This allows recruiters to focus on evaluating and communicating with candidates, rather than manually reviewing each CV.

Internal processes: access, corporate knowledge and reporting

At Headway Inc, an agent was involved in a process that regularly created a load on the platform team — processing requests for access to Google Cloud and Firebase. Agent Gandalf works with requests that come to Jira, Ilya Aladin, Product Engineer at Platform Core, told dev.ua.

“The most frequent request to the platform team is to provide access. In particular, to projects in Google Cloud and Firebase. Such requests are similar to each other, so we passed them on to Agent Gandalf,” he explains.

When a request is assigned to Gandalf, the agent checks if it falls within his area of responsibility, estimates the cost of execution, finds out if the user has the necessary access. He can then explain the rules, ask a clarifying question, forward the request to another team, suggest adding the user to an existing group, or prepare a change in Terraform.

If a configuration change is required, Gandalf creates it and opens a draft merge request. At the same time, he checks whether the proposed access corresponds to the request and whether it does not grant unnecessary permissions. However, the agent cannot independently grant access or close the ticket — this is done by a person. He has read-only access to cloud projects and does not see business data or sensitive information. Before launching, Headway tested the agent on 40 real requests from previous history. In 39 out of 40 cases, he offered the correct or even narrower access than the one that the engineers had provided at the time. The agent never violated the strict restrictions. For comparison: the same model without specially developed skills coped in 80% of cases. The cost of processing one ticket is from $1.7 to $5.2, and the agent infrastructure costs $11.5 per month. The team’s next goal is to make it possible for any division of the company to launch its own agent in one day.

At Sigma Software, agents also work with internal queries and corporate information. According to Yevgen Pak, Senior Support Engineer, they have already become part of daily work, not demonstration projects. “We really automate a lot internally. AI agents are no longer a demo, but part of the team’s daily work,” he notes. Among the tasks are internal assistants who answer questions based on the corporate knowledge base, collect and structure information for teams, help maintain Jira, and prepare status reports and dashboards. Separate tools track team rotation, active access and licenses, analyze logs, and look for anomalies. They can also collect data from various sources, including time tracking information in ETS, and generate summaries for teams.

In recruiting, Sigma Software uses tools for initial screening of candidates, market research, and finding specialists. Another area is the preparation and coordination of documents. The company has separate solutions for this: Contract Harmonization Portal helps work with contracts, CV Generator helps prepare resumes, and Proposal Generator helps create commercial proposals.

“The effect is best seen in routines: what used to take hours of manual work is done much faster by agents, and people focus on the architecture,” adds Pak.

At Ciklum, PRODIGY, an internal AI assistant, helps teams navigate product development methodologies and processes. It takes into account the role of the specialist and the stage of the project to offer appropriate recommendations. This reduces the time employees spend searching for the necessary documentation.

Engineering: from a problem in Notion to a merge request

At Headway Inc, the Alfred agent performs some of the engineering tasks in the Nibble product team. As Dmitry Polurezov, Head of Engineering Nibble, explains, the agent receives tasks in Notion and returns the result in the form of a merge request in GitLab.

Between these two stages, he himself assesses the complexity of the task, writes or changes the code for iOS, Android, Web, Content and Backend, runs tests and security checks. After that, the result is analyzed by a separate reviewer agent who did not participate in the work and can appeal the decisions made.

Alfred creates a merge request explaining the changes and responds to comments from people on GitLab: answering questions, justifying decisions, or making edits. But the final approval and merging of the code always remains with the human.

In the first two weeks, the agent created 29 merge requests across six repositories. Fifteen of these were in Nibble’s product repositories, and another 14 were in the agent’s own repository. A typical task takes about 1 hour and 36 minutes and costs about $30 in tokens.

The company emphasizes that for an agent to work reliably, a powerful model alone is not enough. You need high-quality skills, context, memory, budget, and clearly defined boundaries. That’s why Headway built an AI infrastructure from several layers: from working platforms like Claude, Notion, Jira, GitLab, and Slack to the skills registry, the model environment, and the agents themselves. The next step is to establish interaction between them.

At Railsware, engineers are also experimenting with delegating tasks to agents. Anastasia Vlasova, Full Stack Engineer, talks about the Cyrus framework, which connects Linear with Claude Code.

“We assign a ticket to Cyrus, the webhook transfers the task to an isolated machine, where the Claude Code agent works with a separate clone of the repository and its own API keys, makes changes, and opens a pull request according to our conventions,” Vlasov describes the process.

In Linear, this looks like a normal task assignment: an engineer can open a session, view the agent’s history, and make adjustments to its work. But for now, the team only trusts him with small, isolated changes.

“For more complex tasks, it’s difficult to check the result from the outside. You have to pull the branch and review it locally. […] It works for small tasks, but for large ones it hasn’t been proven yet,” the engineer explains.

Railsware has also automated agent training. After a Claude Code session, a special Debrief skill is launched: it collects feedback and updates the skills and roles of agents. Previously, an engineer had to independently recall the error and add the appropriate rule, but this did not always happen.

“Now you don’t have to correct the same mistake twice,” says Vlasova.

At the same time, the company does not name the exact productivity gain. The engineer gives examples of refactoring a key function and renaming a table throughout the project — tasks that previously required much more effort. But measuring the effect is difficult: as soon as routine work is simplified, the team takes on tasks that were previously postponed.

At Ciklum, AI is also used directly in development: according to the company, assistants generate up to 20% of the code base. It is important not to interpret this figure as the share of code that goes into the product without verification. The generated fragments go through the usual stages of testing, review, and integration for the team, and the responsibility for the final result remains with the developers. That is, AI speeds up some of the work, but does not replace engineering verification.

Content and Marketing: Agent Prepares 70 Pins a Week

Agents can work not only with code and internal systems. At Headway Inc, they are also used in marketing. One example is Pinny, an agent in the brand marketing design team who prepares content for Pinterest. Maria Zavertaylo, Brand Marketing Design Operations, talks about him.

Every Sunday, Pinny prepares posts for the following week: Claude writes the texts, and the OpenAI model generates images. The finished materials are stored in the cloud. On Monday, the agent sends them to the designer for review, and after approval, forms a queue of posts for each day. He reports on the status of the generation and the most successful pins in Slack.

“We use our product content as the basis for Pinterest pins, and Pinny helps adapt it to the platform, experiment with formats, and maintain a regular flow of new posts,” explains Zavertaylo.

Previously, preparing a week’s worth of content took over two hours, now it takes about five minutes. The agent prepares 70 pins per week and an additional content bank. Generation costs about $15 per week. Since the materials are created in advance, the team uses batch processing via the Batch API — it’s half the price.

But Pinney doesn’t define the marketing strategy or decide what content should work for the business goal. He takes on the production side: adapting materials, generating variants, and preparing posts for review. The strategy, creative decisions, and quality control are left to the team.

Another example from Railsware is the work of the content team at TitanApps, one of the company’s products. Viktoriya Golovtseva, Content Manager, talks about using Claude Design to prepare visual materials. The team assembled a design system in the tool based on brand materials from Figma: colors, fonts, components, and design examples.

“We add product context and tasks, get the first version, correct it, and save successful results as samples. For example, I can independently prepare a draft of an infographic for an article, a presentation for a webinar, or a landing page layout to check how the text works in the real arrangement of blocks,” explains Golovtseva.

This allows for faster visual proofing of ideas while you’re still working on the material. At the same time, it’s all about creating drafts and first versions: the final design, product accuracy checks, and editorial decisions are left to humans.

An agent does not mean complete autonomy

In these examples, AI agents perform a variety of tasks: analyzing resumes, preparing code changes, reviewing access requests, creating content, and collecting reports. But almost all processes have one thing in common: a human remains responsible for the outcome.

This is especially noticeable where a mistake can have serious consequences: in hiring, access management, or product development. An agent can prepare a recommendation or even a completed merge request, but does not necessarily have the right to approve it independently. In marketing, he is able to create dozens of materials, but does not determine whether they correspond to the brand strategy. Therefore, the practical value of agents now is not so much in completely removing a person from the process, but in removing repetitive stages of work and freeing up time for decisions that require experience, context, and responsibility. And it is at this level that Ukrainian IT teams are already finding specific applications for them.

“Sorry, I shouldn’t have opened that link.” Ukrainian IT professional tells of a curious and cautionary incident with his AI agent Dot

The National Bank is developing a system of AI agents that simulate the population of Ukraine to assess inflation expectations