Give the agent a real editing surface
Imagine asking a coding agent to build a video editor, then using an agent inside that editor to prepare a first cut. These are two different jobs. A software factory can help construct the application. The finished application also needs deliberate interfaces that let an agent operate it. Both become easier when the editor has a clear model of its work.
Start with sources, tracks, clips, captions, transitions, and export settings. Give each object an identity and defined operations. An agent should be able to inspect the current timeline, propose a change, and receive a structured result. A person should see and edit those same objects in the interface. That shared model connects a request like 'tighten the introduction' to an actual, reviewable edit.
Separate selection from rendering
The expensive creative decision is often selecting the material: where a thought starts, what context a clip needs, where its ending lands. Capture those decisions in an editable clip plan. Include source references, time ranges, transcript spans, caption information, and the reason a segment was selected. The plan becomes a durable artifact that can be inspected before another export runs.
Renderers can then produce different outputs from that plan. A landscape cut, a vertical short, an audio version, and an audiogram can share source selections while applying their own composition rules. A new format becomes a renderer with a defined input. It does not require repeating every upstream analysis step or losing the changes a person already made.
Adapt to the source material
Recordings carry different signals. A performance with an audible audience offers timing cues that a clean studio recording does not. A conversation may need speaker changes and surrounding context. A product demo may depend on the relationship between narration and a visible action. Analysis should identify the available signals and choose an appropriate way to use them.
Preserve enough context for the selection to be questioned. A proposed clip should point back to the original recording and make its boundaries easy to adjust. If an agent selected a moment because it appeared to finish a thought, the reviewer needs to hear that ending and the nearby material. An explanation helps most when it sits beside the artifact it describes.
Use one plan across different levels of delegation
The same editing system can support several ways of working. A person might ask for a set of candidates, approve their boundaries, and then render. They might adjust every clip interactively. For a familiar recurring job, they might let a configured workflow produce previews automatically. All of these paths can read and write the same clip plan.
That makes delegation a practical control. People can change how much work they hand over without moving to a different representation of the project. Keep revisions and make proposed edits inspectable. When an agent changes a timeline, the application should be able to show the change, undo it, and render the resulting state predictably.
Check the mechanics, review the meaning
Automated checks can establish useful properties: source files exist, clip ranges are valid, exports have the requested dimensions, captions fit their timing, and the output is decodable. Reference projects can exercise the render pipeline after changes. These checks give the people reviewing the work a more dependable starting point.
Pacing, clarity, and whether an edit preserves the meaning of a performance still deserve attention in the preview. This is an exploration direction for Rigwright: tools and factory patterns that make ambitious creative applications feasible to build, then give their users meaningful ways to work with agents. The editor is one possible output. The reusable machinery is the larger opportunity.