Post a job and the applications pile up faster than anyone can give them proper attention. So the reading gets hurried. Each resume gets a glance, two reviewers reach two different verdicts, and the criteria written into the role slowly slip out of view. A great candidate can vanish in a skim, and the bar shifts with whoever happens to be on screening duty that day.

Our fix is an agent on Machine that handles the first pass the same way every time. When an application lands, the agent checks it against the role's written rubric, writes up a structured screen backed by quotes, and suggests interview times when the fit is strong. It never turns anyone down. Every yes or no belongs to a person. We use this exact setup on the people who apply to us.

The problem

Inconsistency creeps in at the very first glance. A big pile means quick reads, and quick reads wander. Hand one resume to two reviewers and you can get two answers. Sometimes a candidate is marked down for something the job description never asked for. Good people slip through the cracks, and the standard depends on who is holding the pile.

The common workarounds just move the problem. Keyword filters latch onto the wrong words and silently throw out strong applicants. A hurried human pass loses its consistency somewhere around lunch. And an AI screener judging by its own private sense of "good" creates a fairness problem of its own: nobody can see its reasoning, nobody answers for it, and nobody can audit it.

What we built

With Machine, every new application starts an agent. It gets its own session, which is simply a fresh cloud computer (a sandbox) set up for that one applicant, holding the role's written rubric plus limited access to Google Calendar. The agent compares the resume with the rubric and writes a structured screen listing strengths, gaps and a score, each backed by quotes from the application. If the match is strong, it also suggests interview times on the hiring manager's calendar. Then a person makes the call, every single time.

How it works

The pattern

Each application fires a trigger that opens a session carrying the role's rubric in its skills and memory, plus a narrow connector to Google Calendar. The agent judges by the written rubric alone and shows its evidence. It never rejects anyone by itself, and people make every call.

Guardrails

This agent reads people's applications, and what it writes influences who gets hired. So fairness and human judgment are part of how it is locked down:

• Isolation. Each applicant gets a separate sandbox. The session can see only the resume it was given and can reach only the calendar. All that comes back out is the screen and the suggested times.
• Scoped secrets. Keys for the inbox, ATS and calendar are stored encrypted in the Secrets Manager and only handed to the agents you grant them to.
• Human approval gate. No applicant is ever rejected automatically. Scores come from the written rubric alone, each with its evidence on show, and a person decides whether anyone moves forward.
• Everything is code. The rubric, the agent's setup and its permissions all live as files in the repo. They carry version history, and changing them takes a reviewed change request, not a toggle in some dashboard.

The outcome

The first pass is now steady and easy to check. Each applicant is judged by the same written rubric, with the quotes that earned every score sitting right there, and strong fits come with interview times already suggested. Hiring managers stop skimming and start deciding on evidence. Whether someone moves forward or not is still up to a person.