A doom-loop reliability signal for agentic tool buyers, and a video-review workflow extension for Claude Code.
Doom loops: a named agentic reliability gap (procurement).
When reasoning models get stuck re-analyzing the same steps without progressing, tasks stall and token costs spike. This failure mode, doom looping, doesn't appear in vendor benchmarks: it surfaces in production, during the exact multi-step workflows operators buy agentic tools for. Sam Paech's team at Liquid AI open-sourced Antidoom, an inference-time fix using Final Token Preference Optimization that eliminates doom loops across reasoning model families. For operators evaluating or running agentic AI tools, doom loops are now a named, documented, and addressable failure mode. Vendors pitching agentic products for document processing, research automation, or multi-step workflows who haven't accounted for this have a visible reliability gap in their story.
claude-video: video review inside Claude Code (workflow).
Built by Brad (bradautomates on GitHub), claude-video is a Claude Code plugin that adds a /watch command: point it at any video URL, and it downloads the file, extracts frames at intervals, transcribes the audio, and delivers the full result to Claude in a single step. No separate transcription tool, no frame-extraction pipeline, no manual handoff between tools. Practical for recorded demo review, meeting replay QA, content audits, or any workflow where Claude's analysis on video material fits. Available through the Claude Code marketplace.
Antidoom is available now. The harder question for operators: whether doom-loop rate becomes a standard benchmark column before the next vendor RFP, or stays the failure mode you discover after you've signed. Run the Antidoom eval against your current reasoning-model stack before the next procurement cycle. claude-video is in the Claude Code marketplace today.