The hero video, recorded once and played back

Left: a render of the hero video with real agents, recorded with --record. Right: the same render, with every agent (Obeya's Koordinator and the Tipjar worker) played back from that recording with --replay, on Obeya's live code. Both renders used the Qwen3 voice.

Both videos side by side, in sync: left the recorded run, right the playback. The sound is the playback's.

What it costs

Most of what is left of a played-back render is the narration itself (about 4½ minutes) and staging the scratch Obeya. The 13-minute renders before took longer because the agents took longer than on this run.

The same moments, left recorded, right played back

The agent at work: the same status line, the same 14 steps, the same changed files (+31 −3 in 5 files). The branch and the workspace carry this staging's card id. Only “Open app” is missing on the right: the played-back agent ran no app.

After the two spoken comments: the same reply, the same handover, the same second demo (copied back from the recording), with “Each” in bold.

Later scenes are Obeya's own UI and come out alike anyway.

What you need to know

A playback refuses once the agents would see something else than in the recording: changed instructions, a changed tool, or a tool that refuses now what it accepted then. The render then stops and says why, and the run has to be recorded again with real agents. A changed sentence that the agents hear (the spoken command, the dictation, the comments) also needs a new recorded run, because the recorded agents act on the old words. That case is only logged, not refused.

For feature demos, the run recipe of this repository now says to record the first take with real workers and to play it back for retakes of the same demo.