Expert data · Agent trajectories

Web Agent Trajectory Collection at Production Scale

July 2, 2026

Case study graphic: Biz-Tech Analytics web-agent trajectory collection pipeline showing task volume, review, and quality metrics.
500+
Tasks/Day
10-minute
Per task
90%+
Acceptance Rate
<10%
Client-side Rework

Overview

Biz-Tech Analytics built and ran a sustained web-agent trajectory collection pipeline, generating structured behavioral data for computer-use model training. Each data point pairs a precisely sequenced interaction trace with structured edge-case documentation for deterministic replay and benchmarking.

Annotation Framework

Every click, keystroke, scroll, and URL navigation was treated as an independent, screenshotted, and recorded event. Each step gets a clear label saying what was done and why.

QA Pipeline

Multi-point review at 500 tasks a day

Every task went through a multi-point review before delivery - checking task completion, whether actions were purposeful rather than exploratory, whether the recording was clean, and whether screenshots were usable for replay. This kept client-side rework low (under 10%) while the pipeline held a steady daily output.

Annotators were onboarded to gold-standard delivery within 48 hours, with the active cohort sustaining 500+ long tasks/day across browsers, file managers, spreadsheets and desktop apps.

How we keep an evaluation fair is set out in our independence policy. See also who runs Biz-Tech.

Have a similar challenge?

Tell us what you are building and what has to hold up under test. We will scope the work and say what it takes.