AI agent sent eight job applications in a week but secured no interviews

AI can deliver suitable job openings every morning and draft a tailored letter to each employer. But if it checks only the application status, a recruiter’s reply may go unnoticed. That is exactly what happened to the author of an experiment published on Habr: a message sat in a chat for two days until the person found it themselves. For agent developers, this presents a concrete task: check messages separately from statuses and incorporate any replies into the overall application pipeline.

In the second week, the author used the agent to send eight applications, received two rejections and secured no interview invitations. Over two weeks, the author counted 12 applications. This is a small personal experiment, so it is too early to draw conclusions about AI’s effectiveness in recruitment. But the errors already reveal where automation needs engineering safeguards.

The shortlist is ready, but results have yet to follow

Every weekday, the agent compiles a digest of vacancies, assesses how well they match the author’s experience and explains any mismatches. It prepares a separate cover letter for each application. The author prefers to flag gaps in experience upfront, such as having no experience with ML or high-load systems.

Selection uses a ten-point scale with a threshold of 6.5. It takes into account leadership responsibilities, location, work arrangements, company type and management experience. Agencies and custom software development firms are excluded in line with the applicant’s preferences. If a required skill is missing, the score is lowered or the vacancy is discarded.

However, this score only helps sort job descriptions. The author rejected one vacancy that had cleared the threshold after speaking to an acquaintance who had worked at the company and described poor internal conditions. The agent had no access to that information. The score’s practical value is limited by its input data, and a figure of 6.5 does not turn a job description into an employer background check.

Errors emerge between individual actions

The most telling failure involved tracking replies. An employer sent a standard message saying the resume was under review, while the application list still showed the status “Viewed.” The agent did not record the message in the pipeline. The automation read one source of state information and missed another.

Imagine a typical workflow: in the morning, the agent reports that there are no replies and suggests sending the next applications. In reality, a message is already waiting in an adjacent section. For the report to be useful, the application list needs to be cross-checked against correspondence, with messages linked to specific vacancies. This is a separate integration task that is easy to overlook when building an initial prototype.

Another error affected the log: twice, the agent replaced the file’s contents instead of appending an entry. An “append only” rule was introduced afterward. The engineering takeaway is to restrict the available write operation: give the agent an operation for appending an event, retain history and verify that earlier entries remain intact. An instruction in the context alone had already failed to prevent the error from recurring.

A similar failure occurred when publishing to LinkedIn: according to the author’s account, attaching an image reset the editor, and the post went live without text. This calls for checking the publication’s final state after all attachment-related actions are complete.

What to take back to your product

The agent has delivered something useful: the author no longer searches for vacancies manually, and interview preparation materials are gathered in one place. Yet, by the author’s own account, that prepared resource has never been opened. Creating a file does not mean the next step in the process has happened.

For a product manager, this is a reason to separate the metrics: vacancies found, applications sent, replies detected, invitations received and actions taken by the user. The report also contains an inconsistency: the author mentions three applications in the first week and eight in the second, yet gives a total of 12. That pipeline should be reconciled before drawing conclusions.

For now, the case demonstrates the agent’s usefulness in organizing a job search while also showing the cost of oversight. The developer’s next step is to verify log integrity, comprehensive reading of correspondence and the outcomes of external actions. These checks determine whether an agent can be trusted with a daily routine.