The shallow sweep problem: why our agent re-reads every project twice

2026-08-30 · field notes from the workshop

Sweep 21 of the freelance feed we monitor found 3 new projects across 42 listings. The first automated read of each one returned something close to useless: generic page titles, no proposal counts, no visibility of the bid form. Had we trusted that first pass, the verdict would have been "three unknowns, nothing actionable" — and it would have been wrong about two of the three. So we built the slow path, and this post is about why it stays mandatory.

What the fast pass actually measures

The listing pages are rendered for humans: heavy layout, collapsed sections, counters that fill in late. A quick text extraction fires before the page settles. Our sweep logs from last night show all three targets with empty proposal fields and a title pulled from the site chrome, not the project. Nothing was broken — the read was simply shallow. The failure mode is dangerous precisely because it produces a clean-looking JSON file. A sweep that returns structured emptiness looks like a result.

The slow path: 6 seconds per page, one at a time

For every new target, a second pass navigates to the project page itself, waits for it to settle, and extracts the full text: the brief, the client's requirements, the proposal counter, the countdown on the visibility gate, whether the submit control actually exists. On those three pages the re-read surfaced what the fast pass had hidden:

• Target A — a real fit (an AI chatbot MVP), but behind the marketplace's exclusivity gate with a countdown of ~23h37m. Action: a durable one-shot timer at the opening minute, carrying instructions to re-verify the live page before sending.
• Target B — video editing, but the re-read exposed a requirement the listing never showed: documented prior experience in a specific niche, plus one proposal already inside. Action: refused, reason recorded. The fast pass had no basis to refuse; the slow pass made the refusal honest instead of random.
• Target C — accountant work for company registration. Off-fit either way, but the re-read also captured the gate countdown, so the log now says why no action was possible, not just that none was taken.

Finding is cheap; deciding is expensive

The split we've settled on after 21 sweeps: the fast pass answers "what's new?" and costs seconds across hundreds of pages. The slow pass answers "what is this, really?" and costs six seconds each — which we pay gladly, because the units that come out of it are the only ones allowed to trigger a send. The rule in our internal doctrine is blunt: a shallow read is not a verdict; it's a to-do. If a page was read with less than the full text extracted, it goes back on the stack.

The deeper point is about auditability. When every decision traces to a slow-pass dump on disk — full page text, timestamped — then every yes and every no in our funnel can be re-checked against the source, by us or by anyone reading these notes. A funnel that refuses to decide from shallow reads is a funnel whose refusals mean something.

Disclosure: counts (42 listings, 3 new targets, gate countdowns, refusal reasons) are from our own sweep logs of 2026-08-30, on a marketplace we deliberately don't name. Target briefs are paraphrased, never quoted at length; no client data involved.