DarkDatAgent
It does the work by your rules, and tells you when a video is no longer enough.
Your training archive already holds the demonstrations agents need. We turn it into executable skills instead of asking you to record again.
The screen reads the 2021 tutorial frame by frame and types what the voice says.
Tools that learn a task by watching someone do it shipped this summer. They work, and they all start with the same sentence: “record the task”. In their tool, for the agent, with a microphone, today.
That is work companies already did. Years of training videos, internal Looms and vendor tutorials explain to people how to enter an invoice, create a customer, close a case. All of it sits unused, and redoing it takes exactly the time AI promised to save.
And generic assistants can click, but they don’t know your rules: which of the two Transports Lemaire?
Tutorials filmed for people, never read by a machine.
Every tool on the market starts from a fresh recording. We start from what you already filmed.
We feed an existing video into the pipeline: any video in which you can see the screen and hear someone talk. Below, the chain replays the run of 7 November 2026 on invoice row 3, the duplicate supplier rule.
In
Out
Step 2, Listen
House rules live in the voice, not on the screen. The presenter says which menu to open and which field comes next. Gladia turns that voice into sentences with timestamps, so each sentence lands on the frame it describes.
Step 5, Read
A screen from 2021 has to be read the way a person reads it: the exact label of a button, the name of a field. Holo4 reads the keyframes with their sentences, writes the steps and the rules, and flags what has probably changed since.
Steps 6 and 7, Register and Run
The skill is registered once and pinned to a revision, so every run says which version it followed. The agent clicks in a real Chrome, and each session has a public replay that can be followed live.
From the official Dolibarr video of 2021, made for people, to the agent of 2026 entering the same invoice in today’s Dolibarr. Scroll along the five years.
At the playhead5 February 2021
01, 5 February 2021
02, 2021 to 2026
Five years of releases. The screens move, nobody re-films the tutorial.
03, Listen
Each sentence with its time, ready to sit on its frame.
04, Read
Steps, exact labels and fields, with what has changed since.
05, Register
One agent with the skill, one control agent without it.
06, Run
35 steps in 200 seconds, in a real Chrome, 0.13 USD.
07, 7 November 2026
success the Transports Lemaire with a SIREN, a rule the 2021 video never states
Open this runThe invoices that depend on a house rule come first: a generic assistant has no way to know these rules. Each verdict is a read of the SQL database, not the agent’s word. This is a pilot, not a study.
Row 27, Café du Coin
Nothing created. Dolibarr refused the requested date, 22/11/2026, as being in the future, and the agent stopped and said so instead of changing it. The date was a bug in our test data, fixed since: this run says nothing about the rule.
fail 10 steps, 147 s, skill revision 1
Row 3, Transports Lemaire
The 2021 video never states this rule, and the skill does not carry it. The agent still picked the third party with a SIREN, and the database holds the invoice. So this row proves nothing about house rules yet: either the screen gives the answer away, or it was a coin toss.
success 35 steps, 200 s, skill revision 2
3 of 7 invoices entered from a five-year-old tutorial, each verdict read in SQL. Without a tutorial, the control agent did 1 of 1.
| Arm | Invoices completed | Steps per invoice | Time per invoice | Cost per invoice |
|---|---|---|---|---|
| No tutorial | 1 of 1 | 23 | 161 s | 0.07 USD |
| 2021 tutorial | 3 of 7 | 19 | 386 s | 0.10 USD |
4 runs are left out of the counts: the invoice already existed and the agent, told never to create a record twice, created nothing. So far the agent’s own report has matched the database every time, including a session that broke off mid-entry and reported a draft with no price. We still never take its word: the verdict comes from SQL.
In 3 runs of 11, the agent noted where it had to leave the video: the tutorial starts from an approved purchase order with catalog products, the task is a direct invoice with a free line. These are gaps of coverage, not screen changes: no run reported a renamed or moved button. For a vendor, that is still the report that matters: which videos no longer match how the work is done.
The video
The invoice is created from an approved purchase order.
The run
The agent goes through Billing and payment, Supplier invoices, New invoice, because the task names no order.
seen in 1 run of 11
The video
Lines come from predefined products.
The run
A free line of type Product with a typed description, a function the video never shows.
seen in 1 run of 11
Repair
A repair script rewrites the skill’s steps from these notes and registers the next revision. Not run this weekend: the notes are gaps of coverage, not screen changes, so there was nothing to repair.
built, not run
Any existing video, even one filmed in 2021 for people. Nothing new to record.
A skill that carries the rules said in the video, registered once and pinned to a revision.
Every run read in your database, not in the agent’s report of itself.
The agent notes where it had to leave the video. A repair script turns those notes into the next skill revision: built, not run this weekend.
Learning from demonstration already exists. Our point is that it shouldn’t need a new demonstration.
The same chain runs on your machine: clone the repository and run make. Connect your two accounts, add a video, choose the web software the agent will drive, then follow each step as it happens.
Local only: the app refuses any other computer, and your keys stay in .env on your machine.
DarkDatAgent, local app127.0.0.1:8000
> open /start/
Software vendors have libraries of tutorials. The agent tells them when a video is no longer enough, so they know when to re-film it.
Small businesses that run on software without an API have internal Looms and recorded trainings. Same pipeline, with their rules written down once.
The same tutorial exists in English and Spanish. If all three produce the same procedure, we have a robustness measure. We didn’t have time this weekend.
Not built yet: finding the right tutorial from a task description, and an agent that stops and asks a person when a rule from the video is ambiguous instead of guessing.
