Skip to main content

Ways To Run Automations

Automation Agent gives you several ways to get a workflow done. You can begin with a ready-made tool before configuring AI, then connect an Assistant to Automation Agent or configure Workflow AI when the task benefits from it. These are not competing products you must choose between upfront — most people use several of them, depending on the task.

The Short Version​

  • Run a ready-made library script when one already solves your problem. Fastest, most predictable, and often usable without AI setup.
  • Ask your agent live when the task is project-specific, exploratory, or new. Most flexible.
  • Save your own script when a workflow will be repeated. Reviewable, permission-scoped, and launchable without re-explaining anything — with Workflow AI steps where the workflow genuinely needs judgment.
  • Use local or approved models when company policy limits which AI systems may receive project, transcript, or media-derived data.

Automation Agent checks each saved tool's declared requirements before it starts. Permissions remain available when you need to control which project objects or filesystem areas a workflow may change.

1. Ready-Made Library Workflows​

The Library tab ships tested example workflows: marker import/export, rendering utilities, transcript proofreading and translation, and FFmpeg-based finishing tools.

Running one takes a minute: select it under the Examples root, review what it does and whether it is ready, then run it. Some workflows need no AI at all; some use local Runtime Packs; and some call a Workflow AI provider internally for steps that need judgment. The Library shows those distinctions before you load the tool.

Start here when your problem matches something that already exists. See The Script Library.

2. Live Agent Workflow Via MCP​

This model keeps an agent involved while the task runs.

The agent can inspect the active Premiere project, run read-only probes, export transcript or frame context, validate Runtime DSL, execute controlled actions, and adapt the next step based on what it finds.

Use this model for:

  • transcript-driven timeline analysis
  • rough selects from repeated takes
  • highlight or teaser detection
  • chapter marker suggestions from an edited sequence
  • project-specific cleanup where the exact state is unknown
  • workflows that need external notes, scripts, spreadsheets, or other context during execution

This is also where a general-purpose agent shines: it can combine Premiere state with everything else it can do, such as reading production documents or researching terminology online.

Start with Agent Workflow via MCP when you want this mode, and with the Workflow Recipes for ready-to-paste prompts.

3. Your Own Saved Scripts​

When a workflow will be repeated, turn it into a saved script. The AI can help during authoring, but once the script is saved, it runs later without asking a model to solve the same problem again.

Use this model for:

  • project setup, folder and naming conventions
  • repeated import or organization steps
  • marker and export utilities
  • deterministic cleanup tasks
  • workflows that should be shared across a team
  • restricted production machines that cannot use a live AI agent during execution

Start with Create And Save Your Own Scripts when this is the model you want.

Saved Scripts Can Still Think: Workflow AI​

A saved script does not have to be fully deterministic. It can call a Workflow AI provider — Codex CLI, Claude Code CLI, or a local LM Studio model — for exactly the steps that need judgment, while everything else stays fixed.

Compare the two ways to proofread the transcripts of selected clips:

  • Live agent session: the agent first has to work out how to export the transcripts, then proofread them, then work out how to write the corrections back. Every run repeats that discovery work.
  • Saved script with a Workflow AI step: the script already knows how to collect, chunk, and re-import transcripts. The AI only proofreads. The run is faster, burns fewer tokens, behaves the same way every time, and the logic is a reviewable script instead of a fresh improvisation.

As a rough expectation from our own tests: proofreading a 45-minute interview took about 5–6 minutes as a live agent session and about 2.5 minutes with the ready-made library script — same agent (Codex) in both cases. Your numbers will differ with material, model, and machine, but the direction is typical: the saved workflow spends the AI time only on the actual judgment work.

For the role setup, see AI Setup Roles. For ready-made examples of this pattern, see the transcript workflows in The Script Library.

4. Local Or Company-Approved Models​

Automation Agent does not require one fixed AI provider.

The same Premiere automation layer can be used from different agent environments, including Codex, Claude Code, Claude Desktop Chat, LM Studio, and future MCP-capable tools.

This matters when a team has constraints such as:

  • only approved AI providers may receive project data
  • transcripts or client notes cannot leave a controlled environment
  • a local model is required for sensitive material
  • a strong cloud agent may be acceptable for synthetic development projects, but not for production footage

A common pattern is to develop a workflow online with a strong Assistant, then run the saved result offline with LM Studio as local Workflow AI. See Developing Offline Workflows Online and LM Studio Setup.

Permission-Scoped Execution​

The execution model does not replace permissions.

A saved script and a live agent can both be limited to specific Premiere objects and filesystem paths. For example, a workflow can be allowed to create markers on the active sequence without being allowed to move clips, rename project items, or write files outside a chosen folder.

Before running a new workflow, decide:

  • what it may read in Premiere
  • what it may modify in Premiere
  • which folders it may read
  • which folders it may write
  • whether it should stop for review before applying changes

For the details, see Permissions and Safety.

Decision Guide​