Blog
AI agents for video production: working on footage without downloading it

How post-production teams can hand logging, selects and transcript search to AI agents while the footage stays in one shared drive and streams only when it's needed.
AI agents can take on a good share of the clerical work in post-production: logging what was shot, pulling selects from hours of interviews, searching transcripts for a line someone half remembers, and writing tags and notes against each clip.
The obstacle has always been the footage itself. A single shoot can run to several terabytes, far more than an agent in a sandbox or a chat window can hold.
When the footage lives in a shared drive that streams on demand, the agent reads the parts it needs and writes its notes next to the media, and nobody copies a drive to make that happen.
How footage moves around today
For many small production teams, cards are offloaded to a RAID or a portable SSD, a copy goes to a NAS or a sync service, and a shuttle drive travels between the edit suite and whoever is working from home. Our piece on why an SSD shuttle is not a workflow covers how that tends to go once a second editor joins.
Logging sits on top of all this. An assistant editor scrubs through each clip, types notes into a spreadsheet or the bin, and orders transcripts from a separate service. It's slow, and it's the first job to be cut when the schedule slips. That's why so many teams spend their last week asking where that B-roll went.
An agent could help with much of this, but it can only work on files it can reach. With footage on a local RAID, the agent has to run on that machine, and with a sync service that keeps full copies on each computer, every machine running an agent needs room for the whole project. The guide to cloud storage for video editors compares the main options.
An example: a small documentary company
Take an invented production company of four people making a documentary about a fishing town. They have nine shoot days, about forty hours of interviews and a pile of B-roll: boats, the harbor at dawn, the fish market. The director is in one city, the editor in another, and a producer runs the schedule.
All of the footage sits in one Fabric workspace, organized by shoot day. The editor mounts it with the Fabric desktop app, which shows Fabric as a drive on Mac, Windows and Linux. The clips take no space on her laptop and stream when they're opened, so she can edit straight from the cloud drive instead of keeping a copy of the project locally. Keeping footage off the local disk is also one of the fixes we suggest for a full Premiere scratch disk.
The agents work from the same drive. Here is what they do.
What agents do with the footage
Transcribe and search what was said. Fabric transcribes uploaded audio and video, including MP4, MOV and WebM files, on its paid plans, and the transcripts become searchable alongside everything else in the drive. See audio and video transcription for the details. Once the interviews are transcribed, the director can ask an agent to find every moment where someone talks about the storm that closed the harbor. The agent searches the transcripts and writes what it finds into selects/storm.md: the clip name, the interviewee from the shoot-day notes and the quoted line. When the editor opens a result in Fabric, it jumps to the point in the recording where the line is spoken. Searching what's said in your videos shows how that looks for a person.
Log each shoot day. A second agent works through the transcripts and file names for one shoot day and writes a log. The log is a plain Markdown file in a logs folder, one entry per clip, in a format the team agreed on at the start. Tagging, in this setup, means words the agent writes into that file. An entry might look like this:
Because the log is text, people and agents can search it later with the same tools they use for everything else, and the editor can keep it open beside the timeline.
Check files without fetching them. Agents can read part of a large file without downloading the whole thing. For video, an agent can read a segment of a clip or the part of a file where its technical details are stored, enough to note a duration or confirm the right file is in the right folder, without pulling a four-gigabyte interview across the network.
Search the whole library on the server. Through Fabric's MCP server or the CLI, agents can run grep, regular expressions and full-text search across the entire drive. The search runs on Fabric's side, so finding every log entry tagged market across nine shoot days doesn't mean opening each file over the network. We wrote about why that matters in why AI agents give up searching your files.
What stays with the editor
Everything above works from words: transcripts, file names, file details and written notes. That covers a lot of what assistant editors spend their days on and leaves the visual judgment where it was. Focus, light and which take has the better performance are still the editor's calls, recorded in the same logs as the agents' notes. B-roll with no dialogue gets logged from file names and whatever the camera team wrote down on the day, so it pays to keep those notes in the drive with the footage. See B-roll in Fabric.
Several agents, one library
The documentary team runs one agent per shoot day for the first pass, so nine agents may be writing logs at the same time while a tenth builds the selects lists. Two practical things make that safe.
The first is how Fabric handles writes. If two agents add to selects/storm.md at the same moment and one saves a change based on an older version of the file, Fabric refuses that write. The agent re-reads the file and reapplies its change, so neither set of selects is lost and no duplicate file appears. When two AI agents edit the same file walks through the mechanics.
The second is access. With Agent Vault, each agent is invited as a non-human member of the workspace. The logging agents can read the footage folder and write only to logs, and none can delete anything, because deleting is a separate permission nobody granted. Access lasts a single task, a day or a week and then ends on its own. An agent that needs another folder sends a request and waits for someone to approve or deny it. Every read, creation and change is logged, so the producer can see which agent wrote which entry. Why AI agent access should expire explains the reasoning. Agent Vault is on the Pro and Team plans.
The team keeps its logging rules in a CLAUDE.md file at the top of the project: the entry format, the agreed tags, and a note that agents never rename clips. Agents can't change instruction files like CLAUDE.md, AGENTS.md or SKILL.md in Fabric, even with write access, so one confused agent can't rewrite the rules for the rest. If an agent does tidy something it shouldn't have, deleted files go to a trash you can restore from, and earlier versions of files are kept.
The approach is close to the one in running a research project with a hundred agents: many agents, one shared folder, and results that each later job can read as soon as they're written.
People and agents on the same files
The editor, the director and the agents all open the same folders. When the director adds a note to selects/storm.md saying the second Maria quote is the one, the editor sees it straight away, and so does the agent building the rough assembly list. That's the broader argument in your AI agents need a shared drive, not a group chat: agents coordinate through files, and the files need to be in one place everyone can reach.
Client review happens in the same drive. The producer shares a cut with the client as a commenter, and the client leaves timestamped comments, with drawings on the frame where needed, next to the logs and selects. Getting client feedback on a video shows the steps.
Setting it up
Move the footage into a Fabric workspace, organized by shoot day or scene, and install the desktop app on any machine that edits or runs local agents. Connect remote agents through the MCP server, which works with Claude, ChatGPT, Claude Code, Codex, Cursor and VS Code; how AI agents connect to a shared drive compares the routes. Write a short CLAUDE.md describing the log format. Then give each agent its folders in Agent Vault and a time limit.
Fabric prices per person, so the agents don't need their own seats. See the plans for teams for what's included. If the idea of a drive built for agents is new, what is a shared drive for AI agents? is a good place to start.
Frequently asked questions
Can AI agents log video footage?
Yes, from transcripts, file names, file details and the notes your team keeps. The agent writes its log into a text file, one entry per clip, which people and other agents can then search.
Can an AI agent find selects from interview footage?
Once the interviews are transcribed, an agent can search the transcripts for a topic or phrase and write a selects list with the clip name and the quoted line. Opening a result in Fabric jumps to the moment the line is spoken.
Do I need to download footage for an AI agent to work on it?
No. Agents on a computer with the Fabric desktop app read the mounted drive, where files stream when opened. Agents can also read part of a large file, so a check on one clip doesn't fetch the whole thing.
Does Fabric transcribe video files?
Yes, on paid plans. Fabric transcribes uploaded MP4, MOV and WebM video and MP3, WAV and M4A audio, as well as YouTube, TikTok and podcast links. Transcripts are searchable and can be edited.
Can AI agents tell what's in a shot?
This setup works from words: transcripts, file names and written notes. Judging framing, focus or performance stays with the editor, whose notes can sit in the same log as the agents'.
How do I stop an agent from deleting footage?
In Agent Vault, give the agent read access to the footage folder and write access only where it keeps notes, and leave out the permission to delete. If something is deleted anyway, it goes to Fabric's trash and can be restored.
Can several agents log footage at the same time?
Yes. Each can work on its own shoot day in the same drive. If two agents save to the same file at once, Fabric refuses the write based on the older version and the agent reapplies its change, so nothing is overwritten.
Do agents need their own Fabric seats?
No. Fabric prices per person, not per agent.
Related pages
Otras entradas de blog:

Agent memory vs a shared drive

Can AI agents use Google Drive or Dropbox?

What happens when an AI agent deletes the wrong file

Why AI agents give up searching your files

Why AI agent access should expire

Your agent instruction files are a security risk

When two AI agents edit the same file

Ask your files, don't build a dashboard