AI Production Agent

Overview

The AI Production Agent takes your planned storyboard and produces every shot — generating images, evaluating quality, retrying when needed, and optionally producing video from approved frames. You choose how much control to keep.

Set your agent mode from Project Settings. The agent appears in the storyboard toolbar when your project has shots with prompts ready for production.

Guided mode

In Guided mode, the agent works through your storyboard in phases and pauses at key checkpoints for your approval:

1. Plan review — The agent shows you which shots it will produce, the models it selected, and the estimated credit cost. You approve before any credits are spent.
2. Proof shots — A small batch of sample images is generated first. You review them to confirm the visual direction before full production begins.
3. Image production — The agent generates all remaining frames. Each image goes through prompt optimization and quality evaluation. You review the full gallery before approving.
4. Video production — Approved frames are animated into video clips. You review the final output.

At each checkpoint you can approve, provide feedback on specific shots, or stop the session. This gives you full visibility with minimal effort.

Autonomous mode

In Autonomous mode, the agent runs the full pipeline on its own. Before starting, you configure:

• Credit budget — The maximum credits the agent can spend. Production pauses when the budget is reached.
• Quality threshold — The minimum score (1-10) for auto-approving images. Higher values demand closer prompt adherence; lower values approve more freely.

While the agent works, you can watch the live Activity Log, pause at any time, provide feedback on individual shots, or stop the session entirely. The agent processes shots in parallel up to your plan's concurrency limit.

Prompt optimization

Before the first image attempt for each shot, the agent enhances your prompt with:

• Technical photography terms (depth of field, focal length, exposure)
• Specific lighting descriptions (golden hour, rim lighting, soft diffused light)
• Composition guidance (rule of thirds, leading lines, negative space)
• Scene context from your story (mood, setting, character details)

This happens automatically and improves first-attempt quality. Your original prompt is preserved — the enhanced version is used only for generation. If optimization fails for any reason, the original prompt is used so production is never blocked.

Quality evaluation

After each image is generated, AI vision evaluates it on four dimensions:

• Prompt adherence (40%) — Does the image match what was described? Are the right subjects, actions, and details present?
• Composition (30%) — Is the framing, balance, and visual flow strong? Does the shot type match?
• Technical quality (20%) — Sharpness, lighting, color accuracy, and absence of artifacts.
• Mood (10%) — Does the emotional tone match the scene context?

The combined score is compared against your quality threshold. Images that pass are approved automatically. Images that fail receive a refined prompt from the evaluator that addresses the specific issues — and the agent retries with the improved prompt.

Smart retries

Not all failures are the same. The agent handles each type differently:

• Server error or timeout — The problem is infrastructure, not your prompt. The agent retries the same prompt without modification.
• Rate limit — The agent waits for the appropriate delay before retrying.
• Bad request — The model couldn't process the prompt. The agent simplifies the composition and retries with a clearer description.
• Quality retry — The image was generated but didn't meet the quality threshold. The agent retries with the evaluator's refined prompt.

Each shot is limited to 3 total retries and 2 quality retries. After exhausting retries, the best available result is kept and the agent moves on to the next shot.

Safety limits

The agent includes multiple layers of protection to prevent runaway credit usage:

• Credit budget — The agent stops when the budget you set is reached.
• Per-shot retry limit — Maximum 3 retries per shot for generation failures.
• Quality retry limit — Maximum 2 quality-gate retries per shot. After 2, the image is auto-approved.
• Circuit breaker — If 5 shots fail consecutively, the agent pauses and notifies you rather than continuing to spend credits.
• Manual controls — You can pause, resume, or stop the agent at any time from the production panel.

Activity Log

Every action the agent takes is recorded in the Activity Log at the bottom of the production panel. You can see:

• Image and video generations with model and credit cost
• Quality evaluations with scores and verdicts
• Prompt optimizations
• Approvals and retries
• Pauses, resumes, and budget warnings

The log is collapsible and shows the 5 most recent actions by default. Expand it to see the full history for the current session.

Works with Stages too

The same AI Production Agent that powers Project storyboards also works with Stages. When you enable an agent mode on a Stage, the pipeline produces all the visuals your page needs — background image, hero cover, promotional video, and event music — in one automated flow.

Instead of the scene-by-scene approach used in Projects, the Stage pipeline works through four visual types in sequence. The same quality evaluation, smart retries, prompt optimization, and safety limits apply. In Guided mode you review each visual before the next is generated. In Autonomous mode the pipeline runs end-to-end within your credit budget.

Learn more in the Stages & Invitations guide.

Getting started

For Projects

1. Open your project and go to the Storyboard tab.
2. Make sure your shots have prompts (generated from scene breakdowns or written manually).
3. Go to Project Settings and choose your agent mode (Guided or Autonomous).
4. For Autonomous mode, set your credit budget and quality threshold.
5. Click Start Agent in the storyboard toolbar.
6. In Guided mode, approve each phase as the agent pauses. In Autonomous mode, monitor the Activity Log and intervene if needed.

For Stages

1. Open your Stage in the editor.
2. The production panel appears when your Stage has an agent mode configured.
3. Click Start Production to begin. The pipeline generates background, hero, video, and music in order.
4. In Guided mode, review and approve each visual. In Autonomous mode, watch the Activity Log while the pipeline runs.
5. When all visuals are complete, publish your Stage.

Frequently asked questions