I let the agents run BD overnight. Here’s the morning report.

Most business development is not glamorous. It is reading a company’s latest sustainability report, finding the one number that actually matters, and writing five honest sentences about it before someone else does. It is also, crucially, the kind of work that does not need me to be awake.

So we built an overnight runner. While I sleep, three research agents fan out — one on the company, one on the contact’s recent activity, one on the news — and hand their notes to a drafting step that writes a first-pass email in the house style. Dedup, domain verification, and pacing all happen automatically. By breakfast there’s a Slack summary: who got researched, what got drafted, what to approve.

The night it tried to call Saudi Arabia

Autonomy is humbling. Early on, the dialer-list builder “found” phone numbers that turned out to be area code 966 (Saudi Arabia) and 400 (unassigned). It had cheerfully queued them for a US calling campaign. The fix was a one-line allowlist of valid US area codes — but the lesson was bigger: an agent that will happily do the wrong thing at 3 a.m. needs guardrails that fail closed, not open.

Now the rule is simple. Anything outward-facing — an email send, a Slack post, a dial — passes through a check that refuses unless it’s on an explicit allowlist. The agent can research all night. It cannot embarrass you all night.

It started leaving me notes

The unexpectedly delightful part: we gave the agent a small file-based memory. Now it writes itself notes between sessions — “this deploy clobbers the map token, swap the image instead,” “this client’s data lives in a different repo, don’t mix them.” I came back one morning to a tidy index of everything it had learned the day before. It is, somehow, a better note-taker than I am.

None of this replaces judgment. The morning still starts with me reading every draft and killing the weak ones. But the night shift is handled — and the coffee is better when the report is already waiting.