Create a Scene agent with Director¶

Status: Current | Last reviewed: 2026-10-20

Director creates scripted voice, personal assistant and Agent Assist demos. Each chapter can show a customer phone, assistant call or browser, Salesforce console, or Headless workspace. Participant voices, screen interactions, Assist responses and visual cues share one timeline. The finished MP4 uses saved audio and Assist takes.

Scene agents appear only when your deployment administrator enables them with FEATURE_SCENE_AGENTS=true. They are off by default, independently of Analytics. When disabled, their creation option, filters, saved entries and Director links are unavailable. Saved Scenes and exports are retained.

For a personal AI phone app, IVR navigation, browser handoff and Salesforce Messaging Session, use the personal assistant journey.

For channel selection, IVR diversion and a product video inspection, use the Alder Homeworks walkthrough. Its Guided Service screens, website chat widgets, Customer Video portal, Salesforce Video Call and generated portrait video takes are editable in Director. The revised return Scene keeps typed chats silent, shows Ari thinking and sending, and runs narration alongside navigation. A push notification opens a fresh reminder conversation, and independent customer/representative video tracks support the inspection.

Open your Scene¶

1. In Agents, choose New Agent → Scene Agent.
2. Director opens the starter billing demo immediately, without a setup page.
3. Rename Untitled Scene in the header and edit the demo. Changes save automatically; wait for Saved before closing the page. A conflicting edit in another window is reported rather than silently replacing your work.

There is no separate Save button. If an automatic save fails, the header shows Not saved and Retry save. Your edits stay in the editor while you retry.

Scene agents appear alongside other agents. Use Filters → Type → Scene Agent to find them; opening a Scene goes directly to its Director editor. Renaming the Scene also renames its library card. Back to Agents returns to the full Agents list without applying a type filter, saving any pending edits before leaving. Open Details to edit the Scene agent's Tags, then select Save tags. Tags are saved separately from the recording script. A Scene contains one recording. Its address is /agents/<agentId>/director, with no recording ID to manage. Older links with ?demo=… still open and simplify automatically. Duplicate copies the script, saved voice and Assist takes, and uploaded workspace images; Export on the card opens its recordings. Scenes are saved drafts and exported as videos; they do not need a runtime Publish step.

Older demos opened from a voice agent remain available. Save as Scene agent moves a saved demo into the Agents library, preserving its takes and recordings and leaving the source voice agent intact. New demos started from a voice agent also create Scene agents.

Your demos are shared within the current tenant and saved separately from the agent's configuration. Builders can edit and export. Viewers can open and play saved demos and download completed recordings.

The starter is planned for a three-to-five-minute conversation. Generated speech determines its final duration. Demos may be up to ten minutes.

Write the conversation¶

The canvas has a script panel and a recording preview. Choose a Chapter above the script to move between sections of the same recording. Use Add chapter or Chapter options to add, duplicate, reorder or remove a section; edit its title directly above the moments.

Each + between moments inserts at that exact position, including before the first moment. Choose a spoken line, capability, phone or workspace action, or visual cue. Agent Assist chapters label the two main speakers Customer and Representative. On smaller screens, scroll within the insert menu to see all moment types. Edit spoken words in the card, then expand Voice & timing for delivery and takes. Other cards use Edit moment. Settings stay in the script panel so the preview remains visible. The moment's timestamp jumps the preview to that step. Use its … menu to move, duplicate or delete it.

Paste a conversation accepts Caller: and Agent: turns. A line beginning with # starts a new chapter. Imported chapters are appended to the current Scene.

Choose Insert detail beneath a spoken line to insert a reusable value at the cursor. The menu shows each detail and its current value. The Spoken line preview beneath it shows the resolved words. For example, Your balance is {{balance}}. becomes Your balance is $84.50. when that is the saved balance. The same menu is available for message text, activity results and component text fields. You can also type {{customer}}, {{email}} or another detail directly, including in an entered field value.

Open Details → Reusable details → Add detail, enter a name and value, then choose Create detail. Spaces in names become underscores: account number is inserted as {{account_number}}. The token appears beside each saved detail. Changing a value updates every reference; regenerate only the affected voice takes. Missing or empty referenced details must be filled before export.

Undo and Redo apply to edits made since opening the current demo.

Choose a chapter stage¶

Open Stage, or select the stage summary above the chapter’s script.

Each chapter has its own stage. Follow Salesforce and Headless Scene stages to configure branding, native record layouts, incoming work, human calls and manual Assist cards, button responses, toasts and saved agent output. Voices, visual cues, timing and export are shared below.

Move smoothly between chapters¶

Chapters already act as separate parts of your Scene, each with its own stage. Select a chapter after the first, then open Stage → Transition into this chapter. Choose Cut, Crossfade, Fade through black, or Slide and set a duration from 0 to 3 seconds. Preview transition plays across the boundary. The script and Timeline also show the transition and content start.

The outgoing chapter finishes before the transition begins; the incoming chapter's speech and actions begin afterward. This adds time without trimming dialogue. The outgoing and incoming Scene state stays at its boundary position during the transition. Seeking and MP4 export use the same effect. Use static content at the boundary when a custom component has its own animation clock.

Existing and imported chapters keep their immediate cuts. + Add chapter starts with a 0.6-second crossfade, which you can change or remove with Cut. The first chapter opens directly and has no incoming transition. Cuts between chapters using the same phone or workspace keep that view open, so its cards do not briefly reload at the boundary. Each chapter still applies its own scripted state. Seeking backward restores the state at that time.

To finish with a fade to black, open Details → Ending and set Fade to black (seconds) from 0 to 3. A one-second fade works well for a short ending. The final chapter finishes in full, then fades and holds black briefly. Preview and MP4 export use the same ending. Set the duration to 0 to remove it.

Choose inbound or outbound calls¶

These settings apply to Voice Agent Call stages. Agent Assist call setup is configured in Stage → Incoming work & Omni-Channel for Salesforce or Incoming work & voice tray for Headless.

Open Details → Call setup, or select the call setup card above the first chapter’s script. Call direction is from the agent’s point of view:

Set Ringing before answer (seconds) to control how long this opening stays on screen, up to 30 seconds. It connects automatically before your script begins. The preview jumps to the opening when either setting changes. Use the playback bar to watch it connect.

New starter and blank Scenes use an inbound call with a 1.5-second opening. Allow 1–2 seconds before the first spoken line in newly authored Scenes. Add a Pause at the start when the Scene opens on a desktop stage. Existing Scenes keep their timing until you edit call setup. Changing direction adds a three-second opening if there is none; enter 0 to start already connected. The opening moves all chapters, voices and phone actions together, including moments with advanced timing. Existing voice takes remain usable.

The phone uses the same call screens as live Preview, and its call timer starts when connected. Both settings are saved with the Scene and each MP4 export. Explicit Phone screen moments still let you direct a particular screen later in the conversation.

Choose Scene branding¶

In Details → Appearance → Branding, select an existing template from your tenant. Switch templates to change the logo, avatar, colors and transcript bubbles, or select Studio default to remove the template. Caller and agent display names remain separately editable.

Branding applies to the phone’s web content and AI Activity in both preview and export. Native Messages, Mail, call screens and authentication keep their shared emulator appearance. Published components use the template’s theme colors when they support those tokens; a component with its own fixed colors keeps them.

The Scene saves a copy of the chosen appearance. Later template edits do not change it automatically. Reopen Details and choose Update from template to apply a newer version. Create or edit templates in Tools & Templates → Branding. Existing exports retain their saved appearance.

Add the standard phone ring¶

Choose Insert moment → Sound effect. New sound moments select Standard phone ring, the recorded ringtone used in the Allianz demo. Set its duration and volume, then choose Prepare phone ring. The default 4.5 seconds plays the two-second recording twice with a short pause. Use Advanced timing to place it over an incoming call, ending before the representative accepts. It works without a voice connection. For a chime or other effect, change Sound source to Generate a sound and describe the effect.

Choose voices and takes¶

Open Voices, select an ElevenLabs connection, and choose separate Caller and Agent voices. Search the voice library and use Play sample to audition. Eleven v4 is selected initially; Eleven v3 is also available.

Choose Generate missing takes to prepare the conversation. Generation uses your configured ElevenLabs account. Each line retains its own takes. Select a line’s Voice & timing to supply a delivery direction, retain inline vocal tags, generate another take, listen, and choose the take to use. Stop generating finishes and saves the current take, then stops the batch.

For narration or more speakers, choose Voices → Add persona, enter a name, then select that persona’s ElevenLabs voice. You can add up to 12 personas. The voice dropdown scrolls with the mouse wheel or trackpad; you can also search or use the arrow keys, Home/End and Enter.

Each persona appears in the + insertion menus as a named line and gets its own timeline track. To change who speaks an existing line, open Voice & timing → Speaker. Pasted conversations can use the persona’s name, such as Narrator: Here is how the experience works.

Additional voices play in preview and export. They stay outside the AI Activity transcript by default, which suits narration. Enable Show [name] in AI Activity in Voices when that speaker should appear in the recorded transcript. Renaming a persona keeps their audio; changing their voice requires new takes for their lines. Reassign or remove their lines before removing a persona.

For each persona line, Voice & timing → Who hears this line? defaults to Audience only · narration. Choose Call participants · include in Assist for a specialist or another human on the call. Narration remains audible while the simulated conversation is paused or on hold and is excluded from Assist input. An Assist recommendation is not spoken automatically; write the representative’s reply as its own line.

Editing words, delivery, substituted details or voice selection can make an existing take out of date. Older takes remain available, but export requires a take that matches the current script. Playback never generates speech.

Spoken record bindings such as {{case.status}} use the chapter’s authored record snapshot when generating a take. Workspace actions replay changes to the visible console records dynamically. If a later spoken line should describe a changed value, write that value into the line or use the next chapter’s authored record context.

Direct the phone and AI Activity¶

The phone uses consistent iPhone-style text in playback and MP4 exports: the clock, call timer, caller name and Messages use bundled Inter. Long caller names display in full. These native phone fonts stay consistent when you change Scene branding; your branded cards keep their own typography.

Use + at the desired point in the script to insert:

Messages, Mail, browser chrome, home, lock and call screens use the same phone components as live agent previews. Authentication and Multimodal Connection first show the native incoming-message banner, then the message and browser. Authentication uses the shared simulation screen headed Identity Verification. You can supply message text, sender, email subject and the link shown on the phone. The recorded clock starts at 9:41 and advances with playback: after one minute it shows 9:42. Pausing stops the clock, and scrubbing moves it to the selected time. Messages and AI Activity keep the timestamps of their original events. The same timing applies to exported recordings.

The starter payment form is available without a catalog. Published catalog components are pinned to the selected version. Their nested content is edited with fields and lists. Configure Interactions & responses for authored button results and toasts, or insert Card event / toast for a timed update. For a phone component, choose Card appearance → Slide in to show it arriving. Place a short pause before it when the previous phone screen should remain visible first. The arrival lasts 0.4 seconds and follows playback when seeking. See manual cards and responses.

To record phone interactions:

1. Pause on a visible component and select Record phone interactions below the preview. It becomes available when the phone shows a component.
2. Type, tap a button or scroll inside the phone. Director adds these actions to the script at the displayed playback time. Typing into one field becomes one moment, so individual keystrokes do not clutter the script.
3. Select Done recording interactions. Open a captured moment to give typing or scrolling a duration. Seeking or playing also exits interaction recording.

These controls record actions in the component; navigate other phone screens by inserting a Phone screen moment.

Director stages verification, messages, transactions and AI Activity. These are authored demo outcomes. Use the agent's Preview and Test experiences to exercise its live behavior.

Add narration and visual cues¶

Visual cues work in both Voice Agent Call and Agent Assist stages. Pause at the desired point and choose Add visual cue beneath the preview, then select a visible target. You can also insert Visual cue from a + menu.

Choose a highlight outline, tooltip, spotlight, cursor, or camera focus. Set its target, explanation, color, placement and duration; camera focus also has a zoom control. Pick target in preview changes the target of an existing cue. Targets stay attached to named screen elements as the preview scales. Use Ease in and out for a gradual entrance and exit. A spotlight also supports Background dimming and Zoom to spotlight, keeping the outline, caption and camera move together.

For a Cursor, choose Move from target, set Movement duration, and enable Show a click with its delay. It moves smoothly on a separate layer. Add the matching screen action at the click response time. Keep both targets visible while the cue runs; preview and export use identical cursor timing.

Use Advanced timing to anchor a cue to another moment's start or end, including a screen action. Word anchors are available for spoken lines. To explain something between conversation moments, insert a narrator line between the call participants and anchor the cue to that narration. A missing or hidden target needs correction before the cue can appear.

Choose Inspect at 1:1 beneath the preview to check desktop text and spacing at the recorded resolution. Use Fit preview to return to the scaled view.

Timing and preview¶

The playback bar stays visible. Open Timeline when you need precise timing; its time ruler shows positions within the full recording. It shows only tracks used in the selected chapter. Caller, Agent and named personas contain speech, and Phone contains screen and component actions. Sound effects holds the standard phone ring and generated audio such as chimes. Agent actions holds additional authored system steps shown in AI Activity, and Pauses shows timed pauses. Agent Assist chapters also have Workspace, Agent Assist, and Visual cues tracks. Authentication, connections and components already add their own activity, so you do not need duplicate Agent action moments.

Select a clip to open its settings. Drag from anywhere on the clip to move it; the time readout follows your pointer. Drop to save, or press Escape to cancel. Snap aligns either edge with another clip’s start or end across tracks. Hold Alt, or turn Snap off, to place the clip freely. Overlapping clips remain individually selectable. Clips stay on their speaker or action track.

Use + / − to zoom, Fit to see the chapter, and horizontal scrolling to reach later times. Drag near either side to scroll while moving a clip. You can drop beyond the current ending to extend the chapter. Click the ruler or empty track space to preview a time. With a clip focused, arrow keys move it by 0.1 seconds, or one second while holding Shift.

Moving a clip keeps unrelated clips in place; actions linked to its spoken line follow it. Each completed move is one Undo step. You can also enter a start time in Advanced timing. Clear that start time to restore sequential timing for the moment.

A moment can also follow the start, end or a word in a spoken line. The generated audio's alignment drives word timing. If you remove the referenced word, Director asks you to repair the timing before export.

Use Play chapter, the playback bar, Loop chapter, or the seek slider. Authored taps update the displayed customer component without resetting it. Static forms open with the field values and visible sections authored for that playback position. Seeking backward restores the earlier state. Seeking rebuilds the phone or workspace, entered fields, call state, record updates, saved Assist output and visual cues, and selects the matching chapter. Preview enlarges playback and hides the editor. Escape or Exit preview returns to editing; Space plays or pauses when you are not editing a field.

Details → Show AI Activity is on by default. For a particular phone chapter, use Stage → AI Activity in this chapter → Phone only to center the phone without the panel. Use recording setting inherits the Details choice. Choose inline actions or a separate transcript/actions layout. This setting applies to both preview and export. Activity timestamps show hours and minutes, without seconds. Action durations are omitted from the recorded panel. Under Show in AI Activity on an action or capability moment, enable Show inputs and result in recording to keep that action's details expanded in playback and the MP4.

Export an MP4¶

1. Resolve timing notices and generate all missing audio.
2. Select Export. Director opens Exports immediately, including from enlarged Preview, then saves and snapshots the latest changes.
3. A notification confirms the export has started. Follow progress in its card; you may leave the page while it renders.
4. When ready, select Download MP4.

The recording includes the configured stage and its visible activity or Assist regions. Editor controls are excluded. Voice-only Scenes use 1440 × 1000; Scenes containing a Salesforce or Headless stage use 1920 × 1080. Both use H.264 video and AAC audio, suitable for normal video players and presentation software.

Only one export runs at a time per tenant. Cancel a running export from its card. If the service restarts or an export fails, the script and voice takes remain saved; retry the export. Existing recordings keep their original snapshot when you later edit the demo.