Higgsfield just launched what they call the first fully automated AI agent for video - real shift or just another hype?

Higgsfield dropped Supercomputer yesterday (May 14). It's pitched as one chat that runs research, planning, generation and distribution end-to-end up to several minutes, and user needs just approve what he wants. Spent the evening testing. **The pitch:** The agent plans whatever you told them to do(either it’s a movie trailer or a short clip), picks models from a routing layer (Claude Opus 4.7, Veo 3.1, Kling, Seedance, Nano Banana), executes, and ships. Memory persists across sessions as a visual graph. 30+ connectors (Slack, Drive, Notion, Gmail, Figma). Scheduled tasks via CronJobs. Parallel chats up to 10. **Most surprising part:** It autonomously stitches clips into videos longer than 15 seconds. Sometimes several minutes. Every other agent I've tested bails at the generation handoff or maxes out at single-clip output. Higgsfield claims a 23-minute pilot was produced in 96 hours using this stack, which is consistent with what I saw on shorter tests. **Where it falls short:** Buggy. Just released so expect chats hanging and credit math that doesn't always reconcile. The long-form outputs sometimes slip into AI slop, when you push past 60s the model coherence drops and you get visible drift between segments. I’ve been getting both incredible and bad results. **Why this might actually matter:** Every AI agent until now lived in text and code, Claude, ChatGPT, Cursor, Manus, Operator, they research, code, click around browsers, fill spreadsheets. None of them touched generative content. When you needed a video you opened Sora, Kling or Higgsfield UI, generated manually, downloaded, edited. Whether it's the right execution or not, this is the first time creative production has its own agent category. Anyone else tested it yet or having opinions? Curious what people are getting on multi-minute outputs.