VideoSense · Live replay

Watch a question become a proven answer.

This is a replay of a real session shape: the question is typed, the agent thinks in visible tool steps, then streams a structured answer — and the three clips it returns are real videos from the library, playable right here.

Session replay

VideoSense · skydive library replay
Thinking…

  1. 1
    Solo freefall over the lake  
    High-altitude POV, tracking steady — no other jumper in frame.
  2. 2
    Wingsuiter tracking away  
    Green suit visible against the valley floor, pulling distance.
  3. 3
    Wingsuit close pass  
    Flyby over the reservoir; clip ends before any deployment.
#1Solo freefall over the lake0:08
#2Wingsuiter tracking away0:08
#3Wingsuit close pass0:08
sql_query · phase-tag scan (12 clips)4ms
analyze_video · #7 → retry (frame gap)1.9s
analyze_video · confirm no-canopy ×35.1s
show_video · 3 clips to side-channel0.3s
Ask about your videos…

What makes it read designed

Conclusion first

The answer opens with one bolded headline — the number — then the labelled list. Single-value asks stay a single line.

Timecodes that jump

Every result carries a clickable timecode chip — ▸ 0:04 seeks that clip to the moment. Try them above, after the replay finishes.

Receipts, folded

The tool steps you watched stream by are still there — collapsed into the quiet Steps footer. Expand it to re-read the run.

What's real here: the three clips are genuine segments from the production library (trimmed & compressed for this page), and the interaction shape — typed question, streamed steps, structured answer, quiet footer — is exactly how the product behaves. The replay itself is scripted so it runs without a backend.