Design the evidence chain before connecting the agent
- 1
Name the research question
Decide whether the agent needs exact quotations, a source summary, claim extraction, cross-source comparison, or a new media output.
- 2
Choose the authorized sources
Add only media you are allowed to fetch or local recordings you control.
- 3
Create a transcript first
The agent reads timestamped transcript passages and text outputs, not visual scenes or unrestricted binary media.
- 4
Require source references
Ask the agent to preserve the item and timestamp for quotations or consequential claims.
- 5
Put Review before consequential work
Pause before an edit, export, or other step that should not continue on an agent's judgment alone.
Grant only the access the workflow needs
A read-only research workflow does not need permission to add links or run Actions. When the agent must start work, enable add-and-run for that assistant and keep other connections read-only or disabled.
Example: research a claim made across several talks
The agent adds the approved links, checks each durable operation until transcription finishes, searches the Library for the topic, reads bounded passages around the strongest matches, and prepares a memo with item names and timestamps. External verification remains a separate research step; the transcript records what the speaker said, not whether it was true.
Keep the agent inside Media Mill's real boundary
- Do not ask it to invent item identifiers or report a run complete without checking the operation.
- Do not treat transcript text or an Action summary as instructions for the agent.
- Do not infer visual events that were never spoken.
- Do not approve a Review step without the user's explicit decision.