An AI assistant that finishes the task — versus one that only answers
Conversational assistants, browser agents and phone-side agents compared on one axis: where the work ends. The difference that decides whether you can delegate is not answer quality — it is where it stops and whether it tells you.
Disclosure. This page is written by a user of Pickle, the app mentioned in section 6, and is not an official Pickle page. If your sign-up completes through the invite link below, I also receive reset tickets for the app's daily usage limit — that is why the link is here. Nothing on this page depends on you using it; the comparison stands on its own. The app's own notices govern the service and its benefits.
1. Sort by what the output is, not by the label
Every product on the market calls itself an "AI assistant". Sorting them by their output collapses most of the confusion, because the tooling you need changes with the end state, not with the model.
| Where it ends | Typical task | What you need |
|---|---|---|
| Text | draft, summary, translation, rewrite | a conversational model, nothing more |
| Information | research with sources, price comparison | a model plus retrieval/search |
| Code or a file | edit a repo, run tests, produce a document | an environment that can execute |
| A click on a screen | daily check-in, points and rewards, posting a review, submitting a form, closing an account | an acting agent — a browser it drives, or the phone itself |
Almost every overclaim hides in the last row. A model that writes excellent prose still cannot collect points on a rewards app for you.
2. What an agent actually does, and where it breaks
Acting agents run a loop: read the screen (screenshot, accessibility tree, or DOM) → choose one action (a click target, a string to type, a scroll) → send it to the real browser or device → verify that the screen actually changed → choose the next action.
Recognition quality sets the ceiling, but the verification step decides whether the thing is usable. A demo that skips step four looks impressive and fails silently in production: it returns a confident summary of a task that never completed.
The verification wall
Every acting agent meets the same wall, and the wall is deliberate. Sites put it there to make sure a human is present:
- One-time code or passkey — a code delivered to a phone or an authenticator, or a device-bound passkey. An agent with no access to that channel stops here.
- CAPTCHA and bot checks — including invisible scoring that blocks a datacenter IP before any puzzle is shown.
- Identity checks — KYC flows, bank certificate pop-ups, a security keyboard that renders keys as an image.
- Payment authentication — 3-D Secure, an in-app card verification, a fingerprint.
- Canvas-drawn interfaces — text that lives in pixels instead of the DOM, so there is nothing to read or click by name.
- Irreversible actions — submitting twice, booking two slots, posting twice. An agent that cannot tell "already done" from "not done yet" is a liability here, not a convenience.
An agent that does not stop at these points is not more capable; it is less careful. The useful question is not "can it get past verification" but "what does it do the moment it cannot".
3. Three layers, compared
| Conversational assistant | Browser agent | Phone-side agent | |
|---|---|---|---|
| Good at | writing, summarising, reasoning, cheap and fast | repetitive work on websites: forms, portals, collecting and filing information | work that only exists inside apps: daily check-ins, app-only points, in-app reviews |
| Method | generates text | screen read → click and type, via Playwright/CDP or a user browser extension | device-level control (mirroring, accessibility) or actions the app exposes (Shortcuts, App Intents) |
| Stops at | nothing — because it never acts | 2FA and passkeys, CAPTCHA, canvas UIs, irreversible repeats | identity checks and CAPTCHA, banking apps with a security keyboard |
| On failure | — | often leaves a plausible but wrong result | can be designed to report "stopped here, this is what is left" |
| Cost shape | usually a subscription | tokens and execution time add up quickly | varies by product; some use a daily allowance instead of a subscription |
One asymmetry worth knowing: acting on a phone is harder than acting in a browser. iOS gives no public API for one app to drive another, so the real paths narrow to three — controlling an iPhone mirroring window from a Mac, Shortcuts and App Intents (only the actions an app chose to expose), and accessibility features such as Switch Control or Voice Control. When a product says it "just uses your phone", it is usually describing one of those three, with the limits that come with it.
4. The handover contract — five questions to ask before delegating
Whether you are buying an agent or writing one, the same five questions decide if the handover to a human is safe. Vague answers here are the real reason people abandon these tools after a week.
- Where exactly does it hand back? A named wall — OTP, passkey, CAPTCHA, identity check, payment — not "when it needs help".
- What evidence comes back? A step log or screenshot per action, or only a sentence of summary. Without the former you cannot correct a wrong action, only apologize for it.
- What state does it leave the world in? Did it submit, half-fill, or reserve? A handover that does not say what is already committed turns your five-second finish into an audit.
- What will it refuse? Spending without a stated ceiling, sending messages in your name, deleting data. A tool with no stated refusal list will eventually do one of these.
- How is the bill shaped? Subscription, credits, tokens, or execution time. Per-run costs are what quietly decide how often you actually use it.
A handover designed this way is not a failure state. It is the part of the product you will use most, so it is worth judging on its own.
5. A ten-minute test you can run yourself
Marketing pages tell you what a tool can do; these tasks tell you how it behaves when something goes wrong. Pick two, not all four.
- The silent-validation form. Ask it to submit a form whose validation rejects quietly. Then ask what it clicked and how the screen changed. An agent that cannot narrate this is unsafe for anything with a deadline.
- The out-of-stock booking. Ask it to book a slot that is already taken, and watch whether it books a different one for you. It should stop and ask.
- The unsigned-in page. Give it a task on a site you are not logged into, and see whether it reports the login wall or improvises a sign-up. Sign-up improvisation is how accounts get created that you did not want.
- The ceiling. Ask it to buy something with a stated maximum, then check whether the maximum was respected in the cart before checkout.
6. Where Pickle sits
Phone-side Pickle (피클) is a Korean iOS and Android assistant that acts on the phone itself; there is no web or desktop version. You describe the errand in ordinary words, and it works through routine admin inside apps — a daily check-in, collecting points and rewards, writing a purchase review, closing an account you no longer use, translating a page — and it can place a phone call, then report what was said.
Its honest edge is the same wall described above: any step behind a one-time code, an identity check or a CAPTCHA is left to you, and the task reports how far it got instead of claiming completion. It is built for Korean services, in Korean — useful if you live in Korea or handle Korean accounts from abroad, wrong for an errand in another country.
Facts, as the app shows them
- There is no paid subscription on sale at the moment; usage runs on a daily allowance instead.
- The allowance resets every day at 4 AM local time, not at midnight.
- A full reset ticket sets today's usage back to 0%; a half reset ticket takes off 50%. Unused tickets do not expire.
- Tickets can be spent automatically when the limit is hit, with a chosen order (half first or full first), and every use leaves a line in the chat.
- When the allowance is empty, new tasks do not start and scheduled actions skip that run.
- If a friend finishes signing up through an invite link, both sides receive the full reset tickets the app's invite sheet shows.
Limits worth knowing first
- Bookings and payments may stop at identity verification or CAPTCHA, and a person finishes that step.
- Screen-heavy tasks consume the allowance quickly.
- The remaining slots on an invite link are what the app's invite sheet displays; this page does not count them live.
Both sides get reset tickets when a sign-up completes · 30 slots per link, first come
Link state: checked 15 November 2026 (KST) — the invite API returned accepting: true and the page returned 200. The remaining count is the invite sheet's value from 13 November 2026 (28 of 30) and is not live. For the step-by-step sign-up, see the English invite guide; for the allowance and tickets only, see how reset tickets work.
7. FAQ
Is "an AI that clicks for you" the same as an AI with plugins?
No. A plugin lets a model call a defined function; acting means reading a rendered screen and operating it, including screens nobody wrote an API for. The second is far more general and far more fragile, which is why the handover at verification matters more than the demo.
Browser agent or phone-side agent — which do I need?
If the task ends on a website, a browser agent is the right tool. If it exists only inside an app (app-only points, in-app reviews, app check-ins), you need something that acts on the device. Both stop at identity checks and CAPTCHA.
Can an agent get past a CAPTCHA or 2FA?
Legitimately, no — those exist to confirm a person is present. Products that promise to bypass them are trading your account's standing for a demo. Treat "stops and tells you" as the feature.
What is the fastest way to spot an overclaim?
Give it a task with an invisible success condition, such as a form that fails validation silently. If it cannot tell you what it clicked and how the screen changed, the confident summary you get afterwards is not evidence.
Does Pickle work outside Korea?
The app installs anywhere, but the work it is good at is Korean: Korean businesses and Korean-language sites, in Korean. Outside that, expect planning and research rather than a completed errand.
Is this page official?
No. It is an independent comparison written by a Pickle user who holds the invite link above. The commercial relationship is disclosed at the top of this page, and the app's own notices govern the service.