Key Takeaways

  • Muse took first place on raw video editing by correctly applying brand overlays, setting on-brand captions, and pacing cuts tightly without human guidance.
  • Claude Duet failed the video task immediately because it could not download the source file from Dropbox, proving that basic environment friction still stalls generalist agents.
  • Grok Bot outperformed every competitor in Figma via Model Context Protocol, producing original YouTube-style headline ideas and layouts instead of cookie-cutter templates.
  • Agent performance is strictly split by domain; Muse produced the cleanest video cut but botched Figma design by generating an image of a guest peeking through curtains.

Muse Solves the Video Assembly Line

Most teams waste hours pulling clips, trimming dead air, adding subtitles, and branding video snippets for social distribution. Claire Vo put autonomous agents to work on raw podcast footage to see if computer-use agents could handle the pipeline end to end.

Muse handled the entire video cut cleanly. It did not just trim dead space; it understood branding rules without hand-holding. As Vo observed, “The design of the one that I like the most is Muse. So Muse is very simple. It's got the Howa AI tag in the front. The caption layout is ours. There's no funky stuff. The cuts are punchy.”

Claude Duet choked on step zero. It hit an authentication and download wall and never recovered. “Claude could not do it. Duet could not do it. It would not download a video. So Claude totally faceplanted on this one,” Vo said. When an agent cannot reliably fetch an asset from Dropbox, its underlying reasoning model does not matter. The task is dead before it starts.

Grok Bot Proves That MCP Beats Generic Computer Vision

Moving pixels on a canvas requires precise spatial understanding and programmatic control. When Vo tested the agents on designing podcast thumbnails in Figma, the hierarchy inverted.

Dot and Claude Duet defaulted to sterile, safe templates. Muse failed completely on aesthetic judgment: “Muse did the worst job. Muse pulled our existing thumbnail and then did this really ugly one where he's like Blair is like peeking through creepy curtains,” Vo reported.

Grok Bot ran away with the test. Powered by Model Context Protocol (MCP) integrations with Figma, it bypassed clunky visual guessing and directly manipulated elements on the canvas. “I have consistently tested Grock dot and claude duet on Figma. And Grock is definitely the best at Figma. Grock is like 100% the best at using Figma,” Vo said. Beyond clean layer management, Grok generated fresh creative direction: “Grock actually wins this one in that it gave me really good new ideas for these thumbnails and headlines, ones that we hadn't done that I also think look really nice.”

General-purpose agents often trip over tool-specific APIs. Grok Bot succeeded because direct protocol access gave it the speed and precision needed to iterate on design, rather than pretending to be a human dragging a mouse around a browser window.

What to Do With This

Audit your team's media production pipeline this week and split it into two dedicated agent tracks. Assign video clipping and caption placement to Muse, but pipe your design and thumbnail generation into Grok Bot using Figma's MCP connection. Stop waiting for a single agent to handle both tracks.