Writing to every prospect as if you'd done your homework takes far too long by hand. So most teams quietly drop the homework. They load a template, drop in a first name and hit send on the whole list. It's quick, and it goes straight to the bin. Mail that actually lands needs someone to look into each account and write about what's really going on there, and that is the part nobody has time for.

We gave that part to an agent on Machine, and put a person in front of every send. For each lead it gathers background, writes a first email and a follow-up based on what is really happening at that account, saves every draft in the CRM, and sends nothing until someone approves it, with a cap on how many go out each day. It's how our own outbound runs.

The problem

Outbound forces a choice, and most teams pick the wrong side. Personal messages get answers, but researching and writing them one at a time can't keep up, so the list takes over: one template, one merge field, one big button. You get plenty of volume and very little relevance, and your reply rate says so.

The usual tools don't help much. Sequencers fire the same template out on a timer. AI writing tools turn out smooth prose with no real knowledge of the account, so it sounds personal without being personal. And anything that can bulk-send sits one bad filter away from dropping unchecked email into thousands of inboxes.

What we built

A regular sweep wakes the agent on Machine inside a new session: a fresh cloud computer, or sandbox, with limited access to enrichment, the CRM and email. It takes the batch of new contacts one at a time, digs into each account, drafts a first email and a follow-up sequence based on what it learned, and saves every draft to the CRM. Nothing goes out by itself. Each batch waits for a person to approve it, and a daily cap still applies.

How it works

The pattern

A trigger on a new list starts a clean session with limited connectors into enrichment, the CRM and email. Our outreach approach lives as skills and memory. A person approves every batch and the daily cap paces the sends, so nothing unchecked goes out in bulk.

Guardrails

The agent can write and send email, so sending is the step we hold tightest:

• Isolation. Each sweep gets a brand-new sandbox that is thrown away afterwards. It can reach enrichment, the CRM and email and nothing else. Only drafts come out, and nothing is sent from inside the sandbox. If one contact fails, it's logged and skipped, and the rest of the batch is left untouched.
• Scoped secrets. The enrichment, CRM and email keys are stored encrypted in the Secrets Manager and only handed to the agents you grant them to, when the sandbox starts.
• Human approval gate. Nothing is sent until a person approves it. Each batch is reviewed first, and a daily cap limits volume even after approval, so a bad list can never blast out unchecked mail.
• Everything is code. Configuration, skills and permissions are versioned files in the repo, changed only through a reviewed change request, never a dashboard toggle.

The outcome

Our outbound is now written for each account and sent at a pace we're comfortable with. Every contact gets researched, every email is written around what was found and saved in the CRM, and every batch is approved by a person under a daily cap. Personal outreach can grow because the research grows with it, and nothing unchecked ever goes out in bulk.