A candidate passes the screen, and suddenly booking the next round is a little project of its own. You check the interview plan to see who sits on the panel, hunt for a slot when all of them are genuinely free, offer it to the candidate, and get invites onto every calendar before that slot disappears. No single step is difficult. Multiply it by a few interviewers and a few candidates, though, and it turns into a slog. Each day it waits is a day the loop stands still.

So an agent on Machine handles that coordination daily. For each candidate flagged as ready to schedule in Greenhouse, it reads the interview plan, looks up the panel's free time in Google Calendar, and drafts both a time proposal for the candidate and an invite for each interviewer. It sends nothing and has no say in hiring. A coordinator checks every proposal and invite before anything leaves. That's how our own interview loops keep moving.

The problem

Booking an interview loop means joining up two systems that never talk. The plan sits in Greenhouse: who's on the panel, how many rounds there are and how long each takes. Whether those people are free sits in their calendars. Checking both by hand for every candidate who passes a screen is exactly the sort of admin that falls behind as soon as you have more than one open role.

The usual shortcuts break down. A shared booking link is only as good as each interviewer's habit of keeping it up to date, and it can't tell you who the panel should be for this particular candidate. A coordinator flipping between a stack of calendar tabs and Greenhouse can manage, but not beyond a handful of loops. And a missed invite makes you look disorganised to a candidate who is probably talking to other companies too.

What we built

Once a day a schedule wakes an agent on Machine. It opens a fresh session, a cloud sandbox, and reads every candidate in Greenhouse who is flagged as ready to schedule. For each one it opens the interview plan to find the panel, checks every panelist's Google Calendar for free time, and drafts a time proposal for the candidate plus one calendar invite per interviewer. All of those drafts go to the coordinator. The agent sends nothing and makes no decisions about the candidate.

How it works

The pattern

Every day a cron opens a fresh session with connectors to Greenhouse for the interview plan and Google Calendar for the panel's free time. Our scheduling habits live in skills and memory. The agent drafts the candidate proposal and the invites, then stops. The coordinator confirms and sends.

Guardrails

The agent shapes a candidate's next step in hiring, so the line between "suggested" and "confirmed" is the control that counts:

• Isolation. Each run happens inside its own sealed sandbox. The drafted proposal and invites are the only things that come back out.
• Scoped secrets. Greenhouse and Google Calendar credentials are stored encrypted in the Secrets Manager and only handed to the agents you grant them to.
• Human approval gate. The agent sends neither the candidate proposal nor any calendar invite. Each draft waits for the coordinator to check, tweak and send.
• No hiring decisions, ever. Rejecting a candidate, moving them forward and making an offer are all beyond its reach. It books the next conversation; it never decides whether that conversation should take place.
• Everything is code. Scheduling habits and each system's permissions live as files in the repo. They're versioned and only change through a reviewed change request, not a dashboard toggle.

The outcome

Booking a loop now begins with a proposal and a set of invites already drafted from the panel's real calendars, rather than a coordinator juggling five tabs in search of one time that suits everyone. Their morning starts with a quick review, not a hunt. Every send, and every hiring decision, still belongs to a person.