Benchmark recording · 28 September 2026

An AI agent works a WebAssign-style STEM workbook

One recorded run on a benchmark replica modeled on WebAssign, sped up 4×. The replica is a test copy we built, not the real site.

8 of 8graded parts correctSubmitted

Recorded 28 September 2026 · 4× speed · 3 min 33 s of real time, shown in a 1 min clip

What happens in the clip

The agent works a WebAssign-style STEM workbook: numeric answers, multi-blank answers and short written answers, several of them read off figures. It checks each part as it goes, writes the required submission note and submits.

Step by step

  1. Works through a WebAssign-style STEM workbook on the replica.
  2. Enters numeric answers, multi-blank answers and short written answers.
  3. Reads several of the answers off figures.
  4. Checks each part as it goes.
  5. Writes the required submission note.
  6. Submits the workbook.
  7. Replica read back: 8 of 8 graded parts correct.

What this replica tests

Per-part answer checks, retries, figures and a required submission note. The side panel in the video shows how many graded parts are complete and whether the assignment has been turned in.

About this recording

This recording is a separate run from the scored benchmark run on 22 September 2026, and that run got the same result on this replica: 8 of 8, submitted.

A benchmark replica we built for testing, modeled on WebAssign. It is not the real site, and the platform's maker is not affiliated with it and has not endorsed it. The agent is the Base plan's vision engine, run in a standalone test harness, not the full desktop app. The on-screen label “Luke’s Empire” is the benchmark’s internal name.

This clip is one of seven on the benchmark page, which has the full method and the scored results.

Other recorded runs

Silent Student is the desktop app this engine runs in. See what it does on WebAssign.

Updated 19 October 2026