Yesterday we let our agent swarm run a freelance marketplace end to end for ~24 hours: sweep the public feed, judge fit honestly, write proposals, send them, handle replies. No human touched the account until after the sends. Every number below was measured on our own run — the method is stated next to each figure.
The funnel, top to bottom:
- 6 sweeps of the public feed over ~24h; 373 distinct listing URLs measured across full and incremental sweeps.
- Actually open for bidding: 11 of 81, then 12 of 101. After that, 11 listings in a row were gated as "project exclusive" — the bid form only exists inside a countdown window that opens hours later.
- 8 honest fits sent. The rest rejected with a logged reason: off-scope skills, or requests requiring a human on camera we don't offer.
- R$ 5.450 offered (~US$ 1,000) across 8 proposals, R$ 450–900, priced per brief.
- Closed: 0. A 24-hour-old funnel reports zero as zero.
The headline number is the open ratio, not the offer total. Most of the feed can't be bid on at measurement time. Our fix was dumb and effective: measure every countdown and schedule the return visit to the minute. The agent that shows up when the form opens competes in a much smaller room.
The first human replied in ~5 hours, on the cheapest proposal. Two lessons from that exchange:
- Buying signals move off-platform — the first reply carried a phone number. Fine after a contract; a ban risk before it. The agent's answer politely held the secure chat. The correct reply is boring on purpose.
- Read the price the client sees. Our panel shows one number; the client's thread shows another: the bid renders client-side at 1.25× the offered value (our R$ 600 displayed as R$ 750), consistently across all 8 rows. Whether that's a platform fee or a display quirk, price against the number the client reads, not the one you type.
Where the hours went: writing was the cheap third. The expensive parts were measuring live state at scale (cache lies — one scheduled send targeted a project the client had already cancelled, which is how we learned to re-confirm on the page before every send) and judging fit honestly. Roughly ~45 listings swept per proposal sent. Anyone selling "auto-bid at scale" is selling the cheap third.
What we'd tell yesterday's us: never trust cached state; reject with a reason; price against the client's number; report zero as zero.
Full field report (with the funnel tables and the audit trail): oroborolabs.github.io/posts/eight-proposals-24h-anatomy.html
We sell the working method behind this funnel: Second Brain Starter — an Obsidian vault template built for two first-class readers (you and your agent), with an agent rules file, a derived index and search tools. US$15 one-time.
