VEX · verified execution
a service by Lobster Swamp

Your agent got to step 7 and stopped. We finish it — and send the receipts.

Send us the task your browser agent cannot complete. We run it in a real Chrome, on the real site, through the same interface a person uses — and return the result with a replayable evidence pack. If we cannot finish it, you pay nothing.

Send one task — free See a real evidence pack

First task free. Then $5–9 per completed task. Packs from 50 tasks.

90 seconds, one real run

Recorded on 18 November 2026: a price-and-stock task on a French industrial supplier, plus one task that failed — shown with its failure record, because a pack that only contains successes is not evidence.

Why this exists

74.3%best agent on WebArena, October 2026 — humans: 78.2%
64.4%best published score on WebBench: 5,750 tasks, 452 live sites
1 in 3live-web tasks that never finish, at the top of the leaderboard

Benchmarks are kind: curated sites, fixed layouts, no 2FA. Production is not. The tasks that fail are the ones your user actually needs — an order placed, a form filed, a price checked behind a login. A failed step is not a smaller problem than a missing feature; it is the same support ticket, with your name on it.

Sources: WebArena leaderboard (Steel), WebBench / Skyvern, OpenAI CUA. Numbers move; the gap does not.

How it works

1 · You send the task

URL, goal, and the fields you need back. If the flow needs a login, you give a sandbox account — not your customers' credentials.

2 · We execute

A real browser session, isolated profile, human-paced. Every action is recorded: DOM state, network calls, console, video, Playwright trace.

3 · You get result + proof

result.json in your schema, plus the pack that lets you or your customer replay the run and check every value at its source.

What the evidence pack contains

  • Step screenshots — one per action, named after the action
  • HTTP log — every request and response, with status and timing
  • Console log — page errors included, not filtered
  • Playwright trace — replayable in the trace viewer
  • Screen recording of the session
  • result.json — values, source URL, timestamp, run id

The pack is the product. A JSON without it is an assertion; with it, it is a record.

Price

Task classWhat it coversPer completed task
OpenPublic pages, read-only: price, stock, availability, catalogue, PDF$5
StatefulLogged-in flows, multi-step forms, quotes, orders in your sandbox account, dynamic pricing$9
Pack50 tasks, mixed classes, one invoice, priority queuefrom $250

Turnaround 24–72 h per task, batches daily. You are charged for completed tasks only. A task we could not finish costs nothing and comes back with the failure record, which is often the more useful artefact.

What we will not do

Who runs it

Lobster Swamp is a small engineering lab: browser automation, agent systems, and measurement that survives a second look. The execution side is Playwright/CDP with an evidence recorder; the same harness runs our own pipelines daily. You talk to the person who runs the task, not to a queue.

Send one task

Paste the URL and the fields you need. If it is a flow behind a login, say so and we will send back what a sandbox account has to be able to do. Answer within one working day; the first task is free.