Developing Offline Workflows Online
Offline AI work in Automation Agent is easier to understand when you split it into two jobs:
- Assistant: the AI you talk to while you design the workflow. It helps you figure out the steps, tests them, and fixes mistakes.
- Workflow AI: the model a saved workflow calls later, while the workflow is running.
LM Studio is usually best for the second job. It would be convenient if a local model could also act like Codex or Claude while you build the workflow, but that is usually too much to expect. Local models often do not know enough about Premiere, Automation Agent, transcript handling, validation, and recovery steps to design the whole workflow reliably.
That does not make LM Studio a bad choice. It means the local model should do the part it can handle: a clear, bounded thinking task inside a workflow that was already built.
The Basic Pattern
Use a strong online Assistant, such as Codex or Claude, to build and test the workflow first. Then save that workflow and run it later with LM Studio as the local Workflow AI.
In practice:
- Build the workflow online with Codex or Claude on a test project, duplicated project, or non-sensitive material.
- Tell the Assistant which local model you plan to use in LM Studio.
- Ask the Assistant to turn the working process into an offline workflow.
- The saved workflow keeps the Premiere and Automation Agent steps fixed where possible.
- The AI handoff is sized to what your local model can reliably handle.
You do not need to understand every technical detail. The important instruction is:
Build this as an offline workflow. I want to run the AI part later with LM Studio.
Use the local model only for the parts it can reliably handle.
Keep the Premiere and Automation Agent steps fixed in the workflow where possible.
Example: Proofreading Transcripts
Imagine you want a workflow that proofreads transcripts from selected Premiere project items.
If you ask a strong Assistant such as Codex or Claude to build it, that Assistant can work out the whole process:
- find the selected project items
- check whether transcripts are ready
- export or read the transcript text
- split long transcripts into smaller parts if the local model cannot handle the full text at once
- ask Workflow AI to proofread each part
- check that the returned text is usable
- write the corrected result back or save it where the editor can review it
The local LM Studio model does not need to discover all of that. It only receives a clear task such as:
Proofread this transcript chunk.
Keep the speaker meaning.
Fix obvious transcription mistakes.
Return the corrected text in this exact JSON shape.
That is the key difference. The workflow already knows how to get the transcript, how to split it, how to call the AI, and how to store the result. LM Studio only does the proofreading step.
What The Assistant Should Freeze Into The Workflow
When you build an offline workflow, ask the Assistant to keep all repeatable steps inside the saved workflow. For example:
- how Premiere project items are selected
- how transcript readiness is checked
- how transcript text is exported or read
- how long text is split into chunks
- what prompt is sent to Workflow AI
- what output format is expected
- what happens if the local model returns something invalid
- where reviewed or corrected text is written
As a default, send only the parts that genuinely need AI judgment to Workflow AI. For a transcript workflow, that might be proofreading, classifying, summarizing, or extracting fields. If the target model is unusually capable, the Assistant can choose a larger handoff, but it should be an intentional choice based on that model rather than an assumption.
Tell The Assistant About Your Local Model
Different local models have different limits. Some can handle longer text and structured output. Smaller models may need shorter chunks and simpler instructions.
When you ask Codex or Claude to build the workflow, include what you know:
I want to run this offline with LM Studio.
The model I plan to use is [model name].
Please design the workflow so the local model only gets small, clear tasks.
If needed, split long transcripts into chunks.
If you are not sure which model you will use, say that too. The Assistant should then design for a weaker local model: shorter chunks, stricter output checks, and more review points.
What To Configure
On the authoring machine:
- configure Codex or Claude as the Assistant in AI Setup
- optionally configure LM Studio too, so you can test the local Workflow AI behavior before moving the workflow
- build and test the workflow with representative material
On the offline machine:
- install Automation Agent and the saved workflow
- install LM Studio
- load the local model in LM Studio
- start LM Studio's local server from the Developer tab
- configure LM Studio as the Workflow AI provider in AI Setup
Runtime Packs such as FFmpeg or Python packs are separate from AI setup. If a workflow needs those tools, Automation Agent reports that through workflow preflight and the Runtime Packs tab.