Soulmate agent · analytics
—
—
Where the chats go
Each bar is all chats started in that period. Success rate counts only the chats that set out to make a video, so questions, script ideas and chats still running are left out of it. One chat can deliver several videos, for example a batch of shorts. Click a coloured part of a bar to list those chats below.
Unfinished chats
Every chat in the period that didn't deliver a video, by what happened at its end. Edward checked each of these chats by hand; a model's first read of the transcript filled in the rest. Click a reason to see its chats below.
Over time
Chats started each day (each week for all time), and the finished videos produced from them.
People
Everyone active in the period, with how each of their chats ended. Click a person to list all their chats below.
| Person | Chats | Videos | How chats ended | Done | Partial | Abandoned | Failed | Success | Hours | Spend $ | Main reason chats stopped | Works on |
|---|
Work types
By the skill the agent loaded. "$ per video" is all spend on that work type divided by the videos delivered, failed attempts included. Click a work type to list its chats below.
| Work type | Chats | People | Videos | How chats ended | Success | Renders | Re-renders | Avg min | Spend $ | $ per video |
|---|
Friction
Chats in the period touched by each problem. One chat can have several. Click a problem to list its chats below.
Under the hood
Tool errors, failed jobs and nodes in chats of the period. The agent works around many of these, so this is not the same as the problems people noticed. Click a category to list the chats where it happened.
Every chat
Click a chat to see what happened, why it ended, and a link to it in Soulmate.
| Date | Person | Type | Chat | Outcome | Why it ended | Videos | Renders | Min | $ |
|---|
The dashboard counts every chat with the agent, except admins' test chats and the platform developers' chats (Dmitrii Petrov, Leke Miftari). Admin chats marked as made for production are counted like any other chat. Every chat is downloaded from the agent's API with all its events: messages, agent questions, render approvals, tool calls and errors. A model then reads each chat's condensed transcript and judges the outcome, the number of finished videos, the problems, and why the chat ended. Edward then checked every chat the model counted as unfinished: those where the video was in fact finished count as done, his reason replaces the model's where he gave one, and a few chats he excluded are left out entirely. A chat belongs to the day it started, so the period filter picks chats by their start date. A render is a version of a Seedance, Higgsfield or Kling node; a re-render is another version of the same node. Spend is the cost of every graph run from the chat. Minutes count time between events, leaving out pauses longer than 10 minutes. The data refreshes once a day.