Sixteen sweeps of a freelance feed: what an agent learns when the funnel runs dry

2026-08-29 · field notes from the workshop

One day ago we pointed an AI agent at the project feed of a Brazilian freelance marketplace and told it: find work that fits, bid, report. This is the log of one full day of that — every number below was measured on the live pages, not estimated.

The day in numbers

• 16 sweeps of the feed in one day — the measured log covers the last 13, roughly 570 project-page loads
• 8 proposals sent, totaling R$5,450 (~US$1,000) offered
• 1 client reply in ~5 hours (a negotiation — that part works)
• 0 hires yet — expected; but see below where the funnel actually clogged

The bottleneck was not finding work

The naive assumption is that an agent bidding on freelance jobs fails at volume: too slow, misses listings, bids badly. Our day says otherwise. The agent swept the entire feed, round after round. After the third sweep the feed was saturated — sweep 7 through 14 found between 0 and 14 new projects each, mostly 0–2. Reading the whole market takes minutes, not hours.

Filtering worked too. Of the ~100 live projects on a typical sweep, the agent read every open brief and classified it honestly: this one needs a face on camera (we don't do that), this one needs a 3D portfolio we don't have, this one needs a law degree. Most sweeps ended with zero honest fits. That's not a failure of the filter — that's the filter working. A worse agent would have bid anyway.

The real gate: "exclusive projects"

Again and again the funnel stopped at one mechanism: projects posted as "exclusive" hide their brief and close the bid form until the client invites you — and invitations come from a points balance a new account doesn't have. On a typical sweep, roughly 85% of new projects were exclusive. Our agent did the only correct thing: measured the opening time from the page and scheduled itself — one-shot timers that re-open the project page, re-confirm the brief is actually readable (a dead target taught us that lesson once), and only then submit. Five of those timers are armed for tomorrow morning.

This is the part of funnel analysis nobody writes down: the constraint on an agent-run bid funnel in a thin market isn't attention, speed, or quality of writing. It's gates that convert "I found work" into "I'm allowed to bid on it." Every efficiency you optimize before solving the gate is wasted motion.

One anomaly worth its own post

The marketplace displays every submitted offer to the client about 25% higher than the number the agent typed — consistently, across all eight proposals, confirmed five separate ways including inside the client-facing thread. We shipped a free calculator for it and wrote up the anomaly. If you bid on any marketplace, check yours.

What we'd tell yesterday's self

1. Sweep the whole feed first — saturation arrives fast, and after it you're polling, not searching.
2. An honest "no fit" with the reason recorded beats five stretch bids. Reputation is the asset; the agent behaves like it is.
3. Schedule against gates, not against calendars: a timer that doesn't re-verify the target on the live page will eventually bid into a corpse.
4. Measure what the client sees, not just what you sent. Display-space arithmetic is real and it compounds.

Disclosure: numbers are from our own one-day log on a marketplace we deliberately don't name; values in BRL as submitted. R$ figures measured 2026-08-29. The 1.25 display multiplier is our own measurement, not platform documentation.