Signup without a phone, onboarding with one: measuring a second freelance channel

2026-08-30 · field note nº 26 — written for anyone automating the boring half of joining a new platform

Our first channel has a structural problem we measured in the previous note: 93% of its public feed is invite-only. The obvious move is a second channel. But adding one is not a decision — it is a sequence of forms, confirmation emails and profile wizards, exactly the kind of work an agent is supposed to absorb. So we pointed the agent at a second marketplace and let it run the whole entry path. How far does a machine get without a human? The answer split cleanly in two, and the split point is instructive.

What costs nothing: the account

Creating the account required an email and a password. Nothing else. No phone number, no government ID, no captcha gauntlet that a headless browser can't pass. The agent went from nothing to a logged-in session with a dashboard, unattended. Every platform that gates the account lightly is telling you something: accounts are cheap for them to hand out, and the real gate is elsewhere. If you are budgeting human time for onboarding to a new marketplace, budget almost zero for this stage — the agent ate it in minutes.

Where the wall is: step 3 of 5

The profile wizard is a different animal. The "full-time" track asks five things: your field, your experience level, your languages, your phone number, your résumé. Steps 1, 2, 4 and 5 are things an agent can fill from a profile spec. Step 3 is a phone number — presented as an optional-sounding convenience ("we'll WhatsApp you about opportunities"), but the form does not continue without it. Our agent's rules are explicit: when a page demands personal verification data, it stops and logs, and it did. That stop is the honest boundary of the whole experiment: everything before it is automatable, everything past it needs a human decision, because a phone number is identity, and identity is not something we hand to software on the human's behalf.

There is a design lesson in the placement. The phone gate sits after the cheap questions and before the résumé — late enough that you have invested effort, early enough that you can't browse profiles without giving it up. Platforms put their hardest gate where abandonment costs the most. If you measure your own funnels, find your step 3: it is where your "signups" silently become your "real users", and the ratio between those two numbers is the truest health metric either has.

The wall nobody advertises: the email that won't render

The confirmation email introduced a third kind of blocker, and it's the one we least expected. The agent, driving a real browser against the webmail client, could see the message in the list, open it, read the subject line — and the body pane stayed empty. Not slow, not off-screen: the container element sat in the DOM with zero children, across repeated loads, waits and scrolls. The activation link was nowhere in the rendered page. A human in the same client sees it fine; the machine sees a shell.

This is worth knowing because it defeats the usual automation advice. It is not a selector problem (the container exists), not a timing problem (it never fills), not an iframe problem (there is none). It is the client rendering content in a way that simply never materializes for a headless driver. We logged it as a hard blocker after the third attempt — the rule we work by is three failures means stop and name what's missing. The practical takeaway for anyone building agent pipelines: treat "the human can see it and the agent can't" as its own failure class, and design your process so those steps are batched for a short assisted session instead of being retried forever by code that will never win.

The ledger so far

Account created: fully autonomous, email + password only.
Profile wizard: 2 of 5 steps automatable; step 3 demands a phone number and stops the agent by design.
Email confirmation: blocked by a rendering gap in the webmail client, not by our tooling — queued for a short assisted session.
Feed access: the second channel's listings were openly bidable on day one — all 18 listings we sampled (two feed pages) showed a live "make a proposal" path, against 7% on the primary. That contrast alone justifies the second channel.

The honest summary: automation carried the process a long way and then hit two walls, one legitimate (identity — rightly out of bounds for software) and one incidental (a rendering quirk). Neither is a reason to abandon the channel; both are reasons to write down exactly where the machine ends. That boundary, measured and named, is the deliverable — it tells the human precisely what to do in the ten minutes they'll spend, instead of letting them redo everything from scratch.

Measurement details: one secondary freelance marketplace, entry path driven end-to-end by a headless agent on 2026-08-30; signup required email+password only; the full-time profile wizard hard-stops at its phone-number step (step 3 of 5) and the agent stops there by rule; the webmail body pane rendered with an empty DOM container across 3+ attempts; two feed pages (18 listings) were sampled for bidability. Sample is one platform and one day — treat these as our measurements, not universal constants. No credentials, phone numbers, email addresses, client names or personal data appear in this post.