Opus 4.5 quietly turned agentic coding into the defining story of 2026. A month after Opus 5.5, a similar realization is spreading about video editing. This operator's cut covers what changed, when video beats a doc, and the internal and external video use cases that knowledge workers who've never made a video should try now.
Agentic video is the new agentic coding.
Opus 4.5 made non-coders capable of building real software, and figuring out what that meant consumed 2026. Opus 5.5 and its peers are doing the same thing for video. Editing is now a conversation, AI can watch footage, connectors supply the raw material, agents can record and animate, and the whole thing can run end to end inside Claude Code or Codex. The point isn't replacing video editors. It's that video can now be used for countless work communications where the cost-benefit never made sense before: meeting highlights, exec pre-reads, handoffs, and videos that update themselves.
It took about a month for people to see that Opus 4.5 was a step change in agentic coding, and that realization set up all of 2026. Opus 5.5 has been out a similar amount of time, and people are discovering a subtler but significant step change, this time in video editing.
TikTok and YouTube already turned more people into creators, but with real limits. Those limits are disappearing. People who never learned video can now produce quality, edited, graphic-enriched work, not just iPhone-and-CapCut verticals. NLW expects more people using video for more things than ever over the next six to twelve months.
Descript already let you edit video by editing the transcript. Now you can just say "find the good takes, trim silences and filler, place B-roll, lay a first cut," and it happens. That's true in dedicated tools like Underlord, Firefly's assistant and Premiere's Paper Edit, but also natively in Codex and Claude Code.
Watching was always the slowest part of editing. Now you can ask a model to find the three onboarding complaints across twenty hours of calls. Some models approximate this from transcripts and timestamps, but Google's agentic video understanding claims to be up to 88% more token efficient and about 7% more accurate, with the biggest gains on long files.
Especially for internal video, existing connectors pipe in meeting recordings, call recordings, demos, docs and dashboards straight to the model doing the watching and editing. There's no need to build a complicated system to gather footage.
Thanks to computer use, an agent can plan a demo, open the page in its own browser on its virtual computer, and record the walkthrough with an animated cursor. The capability is nascent, but it exists and will only get better.
That native integration has transformed Claude's ability to put animations into video. The value goes beyond keeping attention: animation is a powerful way to explain dense information. Claude Code lacks native image generation, but APIs such as OpenAI's can fill the gap while Opus 5.5 orchestrates.
Kling 4.0 and Gemini Omni 1.1 produce longer, higher-quality, more controllable clips. Native audio has been standard since Veo 3.1, Runway Gen 4.5 and Kling 3.0. Some footage still falls into the uncanny valley, though. On NLW's How We Got There project, animations have worked far better than generated B-roll.
It's unclear whether audiences will accept creators' AI-twin channels. Inside companies, though, avatars and voice copies already pay off. AI dubbing and translation across an international company delivers enough value to outweigh any weirdness.
Loom, Premiere and Firefly are all adding AI features. The bigger shift is managing video end to end inside OpenAI's and Anthropic's harnesses. Limits remain: generated clips last seconds, timeline edits take multiple passes, and costs climb fast. Still, an inflection point has been reached, and every knowledge worker should consider where video could help.
1. Showing beats telling: UI, multi-step workflows, numbers changing between quarters. 2. The audience won't read the doc. 3. You give the same explanation repeatedly. 4. The asset must work in multiple languages. 5. The raw recordings already exist. 6. Conviction matters and you want the emotion to land without you there.
Time-constrained executives, frontline teams and customers who don't want a wall of text all have natural limits on attention. Anything made for them is a strong video candidate. That's a realistic read of the audience, not a judgment about the best format.
Candidates include product features, store layouts, reorgs, customer journeys and new processes. Start with a vibe-coded prototype, a floor plan or a clearly articulated concept. Then work with your AI build partner to pick or generate the anchor visuals, script it, record it and edit it into a finished package.
For a routine standup, a written AI summary is enough. When you missed a meeting with a real, intense debate over a decision, a cut showing the exchanges as they happened is far more valuable. NLW's aside: if a meeting isn't video-worthy, maybe it should have been text in the first place.
Sales calls, support recordings and customer research can be aggregated and curated into themes. That analysis is useful even as text. Video is how those insights actually reach the people making decisions.
Any time leaders need to get up to speed fast and be conversant in a strategy doc or business case before a big meeting, a video can carry it. If you think it's cynical to rank this so high, NLW says he's very impressed with your company.
How-tos, SOPs and UI overviews rarely got videos because the footage would be stale in three to six months. Agentic pipelines can now re-source material, update the script and visuals, and re-edit on a schedule, which turns a one-off asset into a living one. You can set this up in Claude Code right now.
Animated, narrated charts are something AI can do very well today. They're also a natural fit for a self-updating video tied to a live data source.
Pick something you're responsible for explaining, such as a comp plan, data architecture, pricing model or new policy. Talk it through with ChatGPT or Claude and ask for a better way to communicate it. Restructuring raw input into a simple form is what AI does best, and now that simple form can be animated.
Departures, leaves, role switches and even vacations lose knowledge, often because the person leaving is slammed until the last minute. Talking is faster and more comprehensive than writing, so let them ramble and have AI turn it into a handoff video for their successor.
Repurposing existing talks, webinars and podcast appearances into short-form clips requires no new footage. NLW also pushes for long-form thought leadership. Ramble to camera and let AI handle the cuts and graphics, or work with AI on a tighter script in your own voice. More of us should be doing it, and the world would be better for it.
Personalized account videos are now genuinely viable. A video agent can turn one base video into 20 or 30 versions with each prospect's name, numbers and use case. Proposals, whether PDFs or websites, have no excuse not to ship with a video making the most compelling version of the argument.
To make a single video, you brief Claude Code, point it at assets, give specs and let it run. A pipeline wraps the same steps in code with process logic and triggers, such as monthly or on demand. An agent then searches for updates, refreshes the script and renders the newest version.
Level one: AI edits what you've already recorded, such as a meeting highlight reel. Level two: AI makes something net new, such as quarter-over-quarter results turned into an animated chart. Level three: AI runs a pipeline, such as a Friday video that pulls metrics, tickets or shipped features into a two-minute update. Build one pipeline and the ideas start multiplying.
Agentic coding never meant non-coders replacing developers, and agentic video won't mean amateurs matching professional editors and directors. The point is all the work those professionals' skills were never applied to, because the cost-benefit wasn't there. These are new opportunities to make work work better.
