Pipelines rarely fall apart loudly. They go soft, a little at a time. Seven days pass on a deal with no call and no email, and no one clocks it. A rep promises a close date out loud but never types it in. Something parks in "Proposal Sent" for a month longer than it should. A follow-up task hits its due date and drifts past without a word. Your stage-by-stage report shows none of this. The deal is open, it's counted, and on paper all is well.

That's the gap an agent on Machine now covers for us. Each day it goes through HubSpot looking for those exact signs: deals that have gone silent, blank next steps and close dates, deals frozen in one stage, and tasks that are late. It pings the rep who owns each one in Slack, and the day's worst cases go up to that rep's manager. It mostly reads. It flags, it nudges, and it leaves a deal's stage, owner and amount exactly as they were.

The problem

Don't confuse this with messy CRM records. Two copies of one contact, or a blank job title, is about data quality. A deal that hasn't budged in three weeks is about process: the pipeline has stopped doing its one job, which is moving deals toward a signature. A report organised by stage can't separate the deal that's pushing forward from the one that's been left in a corner. They look identical. Both are open, both sit in a stage, both count toward the forecast.

Most teams lean on a weekly pipeline review, with a manager scrolling through each open deal hunting for the ones that went cold. By then a deal that went quiet on Monday is already a week behind, and the meeting only covers what someone thinks to bring up. Reps are busy selling, not checking their own records. A blank next step or close date tends to sit unnoticed until forecast week, when putting it right has become expensive.

What we built

Once a day, a cron starts an agent in a brand new, sealed-off session that can read HubSpot and write very little back. It gathers every open deal and runs each through four checks. Has there been no activity logged within the set window? Is the next step or close date blank? Has the stage stayed put past the set window? Is there a task on the deal that's overdue? Anything that fails gets flagged. Each owning rep gets a Slack ping listing what needs fixing on their deals, and the worst cases of the day go, separately, to the sales manager's channel. The stage, the owner and the amount on a deal are never changed.

How it works

The pattern

Once a day a cron opens a session whose connector into HubSpot mostly reads. A skill holds the hygiene rules. The agent flags deals that are silent, half-filled, stuck or overdue, pings the rep, sends the worst to the manager, and leaves the stage, owner and amount alone.

Guardrails

An agent that reads every open deal each day has to be trusted, so its reach is small and, where it counts, one-way:

• Kept apart. Each run gets its own sealed-off sandbox. The hygiene flag and the Slack messages are the only things it sends back out.
• Locked-down keys. The HubSpot credential is stored encrypted in the Secrets Manager, added to the sandbox at run time, and only handed to the agents you grant them to.
• Mostly reading. The one thing it writes to HubSpot is an internal hygiene flag. Stage, owner and amount are off limits. Those choices belong to the rep and the manager.
• Pings and flags, that's all. What it produces is a Slack message or a flag. It never acts for a rep.
• All of it lives in code. The stale window, the stage-stall window and the rules for what gets escalated are files in the repo. They're versioned, and you change them with a reviewed change request, not a setting on a dashboard.

The outcome

Deals that once hid until the next pipeline review now show up the very day they go cold. The rep who owns them hears about it, and the ones a manager ought to be watching land in front of the manager too. The agent only flags and pings. What moves, and when, is still up to the rep and the manager.