Agents

An agent is an AI worker with one job, a fixed set of instructions and a fixed set of tools (things it is allowed to do besides writing text). You put an agent into a workflow with an Agent step (or a ReviewLoop or ForEach step); the room steps use several agents together.

In this page, tools are described in plain words:

• Read project files — look at, search and read the files uploaded to the project.
• Write project files — save a new file into the project.
• Write code — work in a private, sealed-off workspace (the "sandbox"): create and edit Remotion code, install packages, and check that the code builds.
• Render Remotion — turn Remotion code into a real video file and save it.
• Stop the run — end the whole workflow with an explanation when it cannot continue.

Can render Remotion means the agent can produce a video file from Remotion code itself. Only three agents can: AuthorAgent, MotionGraphicsPlanner and MotionGraphicsDirector. No other built-in agent and no custom agent can. There is no separate "render" step type: rendering always happens inside one of these agents' steps.

How the Agents page shows each agent

The Agents page lists every agent by what it does for your video — for example "Reads how your app is built" with the line "Learns your app's screens and how they fit together." — grouped under headings such as Understanding your app and Editing your footage. The same plain job names appear on workflow steps and on the run page while a step works. A custom agent keeps the name you gave it.

The internal details — the agent's internal name (such as CodeStructureAnalyzer, the name this guide uses below), its type, its tools and the shape of what it returns — are under Details on each card, and under Technical details on an agent's own page. You do not need them to build or run a workflow.

Which agents can render Remotion

The complete list of agents that can turn Remotion code into a video:

Agent	What it renders	Where its render goes
AuthorAgent (shown as "Author")	The whole promo video	Becomes the run's finished promo (the step fails if it does not render)
MotionGraphicsPlanner	Small animated graphics (titles, lower-thirds, callouts, transparent background) and the scenes shown inside tracked screen inserts (solid, normal video)	Composited into your footage by the VideoCompile step
MotionGraphicsDirector	The same as MotionGraphicsPlanner, from a GraphicsRoom step's final call	Composited by the VideoCompile step

RemotionComponentTranslator writes Remotion code and checks that it builds, but cannot render. Every other agent, and every custom agent, can neither write code nor render.

Code analysis agents

These five agents read a web app's source code and describe it for the promo pipeline. All five: read project files and stop the run only; can render Remotion: no. They are used in the four promo templates (quick-win-promo, lean-context-promo, cinematic-feature-spotlight, story-arc-launch-trailer), not all of them in every template.

Agent	Job	Answer shape
CodeStructureAnalyzer	Maps folders, modules and entry points.	CodeStructureOutput
DependencyAnalyzer	Lists frameworks and libraries.	DependencyAnalysisOutput
ComponentInventoryAnalyzer	Lists UI components, their inputs and purpose.	ComponentInventoryOutput
RouteAndApiAnalyzer	Lists pages, navigation and API endpoints.	RouteAndApiOutput
StyleAndThemeExtractor	Extracts colours, fonts, spacing and branding.	StyleAndThemeOutput

Because they only read, an identical later run may reuse their results instead of asking the AI again.

Promo production agents

These agents turn the code analysis into a rendered promo video. They are used in all four promo templates.

Agent (display name)	Job	Answer shape	Tools	Can render Remotion
RemotionComponentTranslator	Builds the Remotion project: writes scene code into the sandbox and checks it builds.	RemotionProjectBuildOutput	Read and write project files, write code (no render)	No
AnimationStrategyAgent ("AnimationStrategy")	Plans scene order, timing and transitions.	AnimationStrategyOutput	Read project files, look inside the sandbox	No
ScriptwriterAgent ("Scriptwriter")	Writes the narration and captions for each scene.	ScriptwriterOutput	Read project files, look inside the sandbox	No
DirectorAgent ("Director")	Writes the shot-by-shot direction.	DirectorOutput	Read project files, look inside the sandbox	No
AuthorAgent ("Author")	Assembles everything, renders the final promo and saves it. If a Voiceover step ran, it also places the narration audio into the scenes.	RenderManifestOutput	Read and write project files, write code, render	Yes
ReviewAgent ("Review")	Scores the promo out of 10 and lists improvements; used by ReviewLoop steps.	ReviewOutput	Read project files, look inside the sandbox, check code	No

An AuthorAgent step fails if it finishes without having rendered a video.

Video editing decision agents

These agents work on your filmed footage. They choose from the short ids a VideoAnalyze step offered (shots s0, pauses g0, speech t0, ...) and never write a time, a position or a number of their own. They all read project files and stop the run only; can render Remotion: no.

Agent	Job	Answer shape	Used by
VideoStoryEditor	Decides which shots, pauses and spoken phrases to keep.	VideoEditDecisionOutput	Step 2 of every footage template except video-derush-edit-room; fallback for EditRoom
VideoEditDirector	Moderates an EditRoom panel and writes its final keep decision.	VideoEditDecisionOutput	EditRoom steps
MusicSupervisor	Picks at most one uploaded music track (m0, ...) plus how loud (Quiet, Balanced, Feature), how much it dips under speech (Off, Light, Normal, Heavy) and whether it loops (LoopToFit, PlayOnce).	MusicPlanOutput	video-derush-edit-music
SoundDesigner	Plans sound effects: which uploaded clip (x0, ...) plays when which shot, pause or phrase begins, with timing (OnCut, Lead, Lag) and volume (Subtle, Normal, Strong). No effects is a valid answer.	SfxPlanOutput	video-derush-edit-sfx
Colorist	Picks one colour look for the whole video as words: look (None, Warm, Cool, Filmic, Vibrant, Muted, Mono), strength, shadow and highlight words.	ColorGradePlanOutput	Fallback for ColorGradeRoom; usable as its own step
ColorGradeDirector	Moderates a ColorGradeRoom panel and writes its final look.	ColorGradePlanOutput	ColorGradeRoom steps
ShotDirector	Plans AI-generated clips: purpose (Cutaway, ColdOpen, EndCard), where they go, camera and length words, and a description prompt. Cannot spend money itself.	GeneratedShotPlanOutput	video-derush-edit-broll
NarrationWriter	Writes narration lines for the edit, each tied to a shot, pause or spoken phrase (s0, g0, t0, ...), never to a time. Zero lines is a valid answer.	NarrationPlanOutput	video-derush-edit-voiceover, video-derush-edit-dub
NarrationTranslator	Rewrites a narration plan in the language named in the run's request: same lines, same anchors, translated words.	NarrationPlanOutput	video-derush-edit-dub
PickupPlanner	Names the spoken phrases (t0, ...) the presenter flubbed and the corrected sentence for each. Zero pickups is a valid answer.	PickupPlanOutput	video-derush-edit-pickups
VideoReviewAgent (shown as "VideoReview")	Scores a compiled edit using checks ReelBolt measured (sentences not cut mid-way, smooth joins, pacing, graphics coverage, music headroom, and how loud and how clearly the narration was heard).	VideoReviewOutput	The ReviewLoop step of every footage template

Motion graphics agents (these render)

The two motion-graphics agents add designed graphics and screen inserts to filmed footage. Both have read project files, write code, render Remotion and stop the run (no "write project files"). Can render Remotion: yes.

MotionGraphicsPlanner (answer shape MotionGraphicsPlanOutput). Used by step 3 of video-derush-edit-graphics, and as the fallback for a GraphicsRoom step. Its answer has two lists:

• overlays — titles, lower-thirds, callouts and tags, each on an offered graphics spot (placementId such as p2), with a duration word (Short, Medium, Hold). By default each overlay is a real animation the agent writes and renders itself with a transparent background, and returns as renderedAssetStorageKey; plain text on a box is the fallback.
• inserts — tracked screen inserts: for each offered green-screen region (regionId such as r0), a scene the agent writes and renders itself as a normal, solid video, returned as renderedAssetStorageKey, plus a reason. The VideoCompile step then pins that scene into the moving screen frame by frame. An insert must have a rendered scene from the same run; there is no text fallback.

MotionGraphicsDirector (same answer shape). Moderates a GraphicsRoom panel. During the panel discussion it is limited to reading; in its final call it writes the plan and renders the graphics and insert scenes, exactly like the planner.

Tell these agents what to show through the run's brief (the Your brief box when you click Run), for example "In the phone screen, show our app's onboarding screen with the logo animating in". See Tracked screen insert recipe.

These agents are the slowest in the video pipeline, because writing, checking and rendering code takes many steps. They are never reused from an earlier run.

Placeholder agents (not AI)

Two built-in agents exist only to fill the agent slot of non-AI steps and are never run as an AI:

• ExtractTransform — used by Extract steps.
• VideoTransform — used by VideoAnalyze, VideoCompile, VideoGenerate, Voiceover, EditRoom, GraphicsRoom and ColorGradeRoom steps. For the three room steps, the real AI panel members come from the room's settings, not from this slot.

They have no tools. Can render Remotion: no.

Custom agents

A custom agent is one you create yourself on the Agents page: you give it a name, a description and a system prompt (its standing instructions). Built-in agents are read-only; custom agents can be edited and deleted.

Important limitation in this version: custom agents do not run inside workflows yet. You can create, edit and add a custom agent to an Agent, ForEach or Parallel step, but the workflow engine has no runner for custom agents, so when the run reaches that step it is marked Skipped ("No agent registered ...") and the run moves on to the next step without doing anything. Use a built-in agent for any work you need done, and steer it with the run's request text (see below).

What is already stored for a custom agent, ready for when they can run:

• Skills: you can attach Remotion knowledge packs ("skills") to it — remotion-create, remotion-markup, remotion-render, remotion-captions, remotion-multimedia. A skill is reference knowledge, not a tool, and does not grant any new ability. (Because custom agents do not run yet, attached skills currently have no effect.)
• Which AI model it uses: an admin can give any agent, built-in or custom, its own AI provider (the per-agent provider override on the agent's page). Otherwise the platform's default chat provider is used. For built-in agents this override is fully active today.
• Tools it would get: only "read project files" and "stop the run". A custom agent can never write code, write project files or render Remotion, whatever its prompt says.
• Answer shape: free-form text, so even once custom agents run they cannot stand in for a built-in agent whose answer another step reads (for example a compile step's decision or graphicsPlan).

Changing a built-in agent's behaviour

You cannot edit a built-in agent's prompt from the Agents page. Ways to steer one:

• Give instructions in the run's brief (the Your brief box that opens when you click Run). Every AI agent in the run reads it.
• Choose a different context mode on its step (which earlier results it sees).
• Ask an admin to assign it a different AI provider.
• An operator can override any built-in agent's prompt in the server configuration (Agents:<AgentName>:SystemPrompt).