General Video Agent wants to automate editing without locking creators into an MP4 file.

Motion opens the beta for General Video Agent, an AI agent capable of editing, generating, and modifying videos exportable to Premiere Pro or After Effects.

AI video editing is no longer limited to generating a few shots or automatically cutting a sequence. With General Video Agent (GVA), Mosaic AI Labs offers an agent capable of handling an entire production, from B-roll analysis to editing, motion graphics, and audio sync. Integrated into its Motion platform, the system combines generative models, computer vision, video programming, and multimodal language models. Its distinctive feature lies in the nature of the produced projects: videos can be corrected in a timeline, modified directly in the preview, and exported to software like Adobe Premiere Pro, After Effects, or DaVinci Resolve. The public beta is now accessible without a waitlist. Editing existing footage or creating new shots

General Video Agent aims to unite two activities that are still often separate in AI tools: sequence generation and the editing of already shot content. Users can submit a brief accompanied by B-roll, style references, or graphic assets. The agent then analyzes the available material, organizes the video's flow, and selects the operations required for its creation. It can also produce new visual assets when the project calls for it, rather than working exclusively with imported files.

The use cases presented by Mosaic notably cover launch videos, product demos, interviews, multicam podcasts, and social media clips. An interview can be re-edited with reframing, animated titles, and B-roll sequences, while an ad brief can turn into a video combining motion design, voiceover, and music. The agent also handles more traditional post-production operations, such as color grading, audio editing, or track synchronization.

To guide the results, Motion can analyze a reference video and extract certain filmmaking principles: editing pacing, typography, transitions, and shot composition. This capability does not guarantee the exact reproduction of an art direction, but allows visual examples to be used as production guidelines alongside text. A project that you can keep editing

One of GVA's main selling points concerns modifying the results. Many video generators produce a finalized file whose internal elements are difficult to access. Any composition or typography correction might then require a complete re-generation. Motion aims to maintain an editable representation of the video, complete with its sequences, graphic elements, and audio tracks.

Users can request a modification in writing, select an element directly in the preview, move a title, or draw an annotation on the image to highlight the area to be corrected. A native timeline also allows users to move sequences, adjust their duration, or modify certain elements without systematically relying on a new prompt. The conversation with the agent thus becomes a way to steer the edit without eliminating opportunities for manual intervention.

This logic extends to exports. GVA announces compatibility with Premiere Pro, After Effects, DaVinci Resolve, Remotion, and Hyperframes, so projects can be picked up in professional or programmatic environments. Mosaic specifies, however, that preserved features vary depending on the destination software. An exportable project does not necessarily mean that all effects, parameters, or behaviors remain identically editable in every application. Multiple technologies behind a single agent

GVA does not rely on a single video generation model. Its operation is described as an assembly of diffusion models, computer vision systems, multimodal models, and programmable video technologies. The agent selects the tools suited for the requested operations, whether analyzing B-roll, producing typographic animation, or building a sequence from code.

Mosaic emphasizes a video representation independent of the framework used to produce it. To study this approach, the team conducted an experiment using the same model, files, and brief across several programming environments, including Remotion, Hyperframes, Python, C++, Forth, and Uiua. Versions produced with languages familiar to the model yielded close results during company-presented evaluations.

In terms of generation costs, Motion shows an average of about $2.86 per video in one of these comparisons, versus $3.12 for Python, $3.13 for Hyperframes, and $3.30 for Remotion. These amounts represent token costs billed by the model provider for the selected trials, not the commercial price of a finished video. The tests remain internal, rely on a limited number of productions, and do not allow for the conclusion that one framework is systematically superior to the others.

The company primarily draws a conclusion regarding the importance of the harness—that is, the set of instructions, tools, planning mechanisms, and checks surrounding the model. Two systems using the same language model can produce noticeably different edits depending on how they organize workflow steps. This is an hypothesis supported by Mosaic's internal experiments, rather than a general law proven for all video architectures. Connected files, applications, and working memory

General Video Agent can access resources hosted across several professional applications. Announced connections include Google Drive, Dropbox, and Frame.io for audiovisual files, as well as Notion, Slack, GitHub, and Figma for documents, communications, and context assets. The agent can leverage this information to prepare a production, retrieve references, or apply brand guidelines.

This dimension goes beyond simple media import. Mosaic presents GVA as a system capable of maintaining a working context and enriching it over interactions. A team that regularly produces videos for the same client could thus reuse graphic assets, editorial references, or information previously associated with its past projects.

Accessing connected services nevertheless assumes appropriate authorizations. The announced capability to read or write to certain sources does not mean all data automatically becomes available, nor that every