How I find customers with a coding agent, day by day

The pitch for agent-run sales usually skips the day-to-day. Here is mine, unskipped. Two sessions a day run on a schedule, and I spend maybe twenty minutes on the whole thing.

Morning: who wrote back

I open autark.sh/autark. The first thing on the page is replies — not hypotheses, not metrics, replies. One row per conversation, the raw thread one click away. Red rows on top are escalations: places the agent decided a human should look.

When someone deserves an answer, I click the row. The lead id lands on my clipboard. Then I go to my terminal:

reply to lead 72716547-... — he asked how the playbook handles
follow-ups. answer that one question, keep it to four lines.

Claude reads the lead (autark lead show), reads the thread (autark mail thread), drafts, runs the draft through autark mail lint — which catches the AI tells, the double questions, the paragraphs nobody asked for — and sends with autark mail reply --lead-id. The send records itself on the lead. I never see a CRM.

That is the entire morning. Most days it is five minutes, because most days two or three people wrote back, not twenty. We send little. On purpose.

Background: the sheet fills itself

At 9:00 and 19:00 a scheduled tick starts an agent session with one job: sourcing. It reads the product context, picks the sharpest active hypothesis, and goes hunting — public posts, repos, launch pages, anywhere the bet says these people show up.

A hypothesis is a frozen bet, two paragraphs: the story (why this product matters to this cohort) and the who (where these people are findable). The current one for Autark is H23 — solo founders who shipped a real product in the last thirty days and just hit the get-first-customers wall.

For each person worth keeping, the agent writes one JSON file — identity, evidence, and an angle that answers "why this human, under this bet" — and records it:

autark lead add --hypothesis-id "$HYP" --run-id "$RUN" --input @/tmp/lead-4.json

Eight to fifteen leads per run, then a narrative: where it looked, what patterns showed up, where the next pass should go. If it cannot write a concrete angle for someone, it skips them. The skill makes skipping the obedient move.

I do not watch this happen. I read the narrative later, like a standup note from a colleague who works nights.

The steering wheel: one feedback note

Here is the part that actually matters, because it is where the loop earns its keep.

Early on, the outreach pitched Autark directly. Forty-seven cold sends, zero replies. The fix was not new software. It was one note, left on the product:

Stop pitching autark in cold outreach. One short personalized line about their moment, then offer the playbook: "i put together a playbook for running cold outreach with agents — just reply yes and i'll send it over."

The agent reads product feedback at the start of every run. Next run, every draft followed the note. Replies started arriving — people saying "yes," which is a strange and wonderful thing to find in a cold-outreach inbox. The whole exchange is sitting in the public ledger if you want to check my math.

That is the management model: you do not review every message, you correct the policy once and the agent applies it everywhere after.

Weekly: kill bets, place bets

Once a week I look at the hypotheses tab. Each card shows what the bet collected and who replied. A bet that produced silence for two runs gets marked dead — not deleted, dead, so the reasoning stays on record. Then I write the next one. After a dozen runs you stop guessing who your customer is; the sheet has been telling you.

What you need to copy this

An agent you already pay for — Claude Code or Codex — and the structure. The structure is the autark CLI plus the skills, and the whole system runs from your laptop, sends from your inbox, and keeps every byte of the data in your account.

The agent does the hours. You do the judgment. That split, it turns out, is the entire product.