Agentic Video Is Real and Here’s What You Should Try

Context

NLW argues that Opus 5.5 has done for video editing what Opus 4.5 did for agentic coding in late 2025. That earlier shift took about a month for people to notice, and it then defined how knowledge workers spent 2026. He believes the same quiet step change is now happening with video, so people who have never edited a video can produce edited, graphic-rich work with AI agents. The episode covers what changed in the models and tools, when video beats a written format, and which internal and external use cases to try first. For PMs, video is one of the main ways ideas, customer evidence and decisions travel through a company, so a sharp drop in the cost of making it changes how product work gets communicated.

The Big Idea

AI agents can now plan, watch, edit and animate video from plain-English instructions, so video becomes a normal work format instead of a specialist one.

The real gain is not amateurs matching professional editors. It is that video can now be used for the many work tasks where the cost of making it was never worth it before.

Key Insights

1. Editing became a conversation

Video editing used to involve many steps: planning and scripting, gathering b-roll, recording, cutting, fixing sound, adding graphics, translating and making short cut-downs for social channels. That meant either a lot of people or a few people working serious overtime.

• Descript had already simplified editing by letting people edit a video by editing the words of its transcript.
• Models are now good enough to take plain-English instructions such as "find the good takes, trim silences and filler, place b-roll, lay a first cut."
• This is no longer limited to video tools like Descript's Underlord agent or Adobe's Firefly AI Assistant. It is now a core capability of general agents like Codex and Claude Code.

2. AI can now watch footage

NLW thinks this is the change people underestimate most, because watching was always the slowest part of editing.

• You can now ask a model to "find the three moments where customers complained about onboarding across 20 hours of calls."
• Some models still work from the transcript and timestamps to approximate watching, but that is starting to change.
• Google launched agentic video understanding in early September, saying it is up to 88% more token efficient and about 7% more accurate, with the biggest gains on long files like 90-minute lectures.

Watching is what lets an editing request actually get delivered, because the agent can find the moments you describe.

3. Raw material and recording are automated

Two further changes mean the agent can gather and even create its own source footage:

• Connectors: existing connectors to meeting recordings, call recordings, demos, docs and dashboards feed raw material straight into the model, which matters most for internal videos.
• Computer use: an agent can plan a product demo, open the page in its own virtual browser, and record the demo with an animated cursor. NLW calls this capability early but improving.

4. Animation and generated footage matured

• Claude Code can write animations in React and render them with Remotion agent skills, which NLW says transformed Claude's ability to put animations in videos.
• Animations are not just decoration. They are a strong way to explain complex or dense information.
• Claude Code has no native image generation, but it can call image APIs such as OpenAI's while a model like Opus 5.5 runs the overall project.
• Generated clips from models such as Kling 4.0 and Gemini Omni 1.1 are longer, better and more controllable, and can fill b-roll gaps when stock footage is missing.
• Since Veo 3.1, Runway Gen 4.5 and Kling 3.0, leading video models generate sound along with the picture.

NLW's own experience making his long-form series "How We Got Here" is that animations worked much better than generated b-roll, which can still fall into the uncanny valley.

5. Avatars fit work better than social

Digital-twin video avatars and synthetic voice copies have existed for a while. Some creators run secondary channels made entirely of AI versions of themselves, and NLW says the jury is still out on whether audiences accept that. Inside a company the trade-off is clearer. AI dubbing and translation across an international company is a case where the value of the output clearly outweighs any weirdness in the avatar.

6. The general agents matter most

Tools like Loom, Premiere and Firefly are adding advanced AI video features. NLW thinks the bigger shift is that people can now manage a video end to end inside Codex or Claude Code. In his view, the spread of video making will come from the agent harnesses offered by OpenAI and Anthropic more than from any single video tool.

7. Real limits still apply

• Generated clips are still seconds long, not minutes.
• Timeline editing is imperfect and often needs several passes.
• Advanced animation, generation and editing can get expensive fast.

None of these limits change NLW's view that the average knowledge worker should now spend some time thinking about where video could help their work communication.

Mental Models & Frameworks

Six signs video beats text

NLW offers six conditions under which a video beats the format you would normally use:

• Showing beats telling: the subject benefits from movement or pacing, such as a UI, a multi-step workflow, or numbers changing between Q1 and Q2.
• The audience will not read the doc: time-constrained executives, frontline teams, or customers who do not want a block of text.
• You explain the same thing repeatedly: onboarding and customer education are the obvious examples.
• It must work in several languages: translation used to be very hard and AI has made it incidental.
• The raw footage already exists: nobody will watch a 60-minute Zoom call they missed, but a 90-second or three-minute highlight version is plausible.
• Conviction matters: when you want a decision and the emotion behind it to land, even if you are not there to deliver it yourself.

Use the list as a filter before making any video. NLW stresses the goal is not to turn everything into video.

A video versus a pipeline

NLW draws a sharp line between the two:

• A video: you open Claude Code, explain what the video should say, point it at the assets, give specs like rough length, and let it run.
• A pipeline: you build an application with process logic and triggers, such as the start of every month. An agent spins up, searches for updates, refreshes the script, and combines the latest source material into a new version.

This is what makes the "video that updates itself" possible. Material like how-tos, standard operating procedures and UI overviews used to go out of date in three to six months, so a video rarely justified its cost. A pipeline turns video from a one-time asset into a living one, costing only the setup and the tokens per update. NLW says this can be set up in Claude Code today.

Three levels to get started

1. AI edits what you already recorded: for example, ask the AI to make a highlight reel of a recorded meeting.
2. AI makes something new: for example, feed last quarter's and this quarter's results spreadsheets into Claude and have it build an animated chart sequence, or build a concept video for a product feature idea.
3. AI runs a pipeline: for example, every Friday an agent pulls metrics, tickets or shipped features and renders a two-minute update that is shared through a set process.

NLW says that once you build one pipeline at level three, you will understand the idea much better and start finding new uses for it.

The concept video recipe

For a product, a concept video can start from a vibe-coded prototype. You can walk through it in Loom or let an agent record the walkthrough with computer use. For abstract concepts like a new onboarding flow or a reorg, NLW suggests three steps:

1. Articulate the concept, with AI's help if needed.
2. Discuss with your AI build partner the right way to visualize it.
3. Pick the anchor visual, such as a walkthrough of a flow, a prototype, or a floor plan and mockups for a new store layout.

From there the AI can help generate visuals, script, record, find external assets and edit everything into one package.

Trade-offs & Nuance

Not every meeting deserves video

A written AI summary is often enough, especially for routine stand-ups with little new information or debate. A short video earns its place when there was a meaningful or intense debate about a decision, because a viewer sees the real exchanges, not just a transcript. NLW adds that if a meeting is not worth a video, it may not have needed to be a meeting at all.

Amateurs will not match professionals

NLW is clear that non-editors using these tools will not produce work as good as real editors, producers and directors using the same tools. He compares it to agentic coding, which was never about non-coders replacing developers. The point is that professional video skills were never going to be applied to these everyday work tasks, because the cost-benefit ratio was not there.

Generated footage versus animation

Generated b-roll can look uncanny, and in NLW's own production animations worked much better. He still expects plenty of settings where generated footage that is not animation will be useful, so the choice depends on the job.

Practical Application

Turn customer calls into a video

If your company records sales calls, support calls or customer research, have an AI aggregate the recordings and pull out the main themes. Then turn the strongest moments into a short video for decision makers. NLW's point is that customer insights often go nowhere unless they reach the people who decide, and a video of real customers saying it is more likely to get there.

Ship concept videos, not just specs

The next time you pitch a feature that does not exist yet, build a vibe-coded prototype and have an agent record a walkthrough of it. For a non-product change like a new onboarding flow, agree with the AI on how to visualize it before scripting. Video carries far more information than a written spec for things like flows and customer journeys.

Make an exec pre-read video

Before an important leadership meeting, turn the strategy document or business case into a short video so executives arrive able to discuss it. NLW frames this as a realistic response to the fact that leaders often do not read the doc beforehand.

Record a spoken handoff

When you change roles, go on leave or take a vacation, talk through what your successor needs to know and let AI turn it into a handoff video. NLW's reasoning is that talking is faster and more complete than writing, and it does not require well-structured thoughts, which matters when people are busy right up to their last day.

Narrate the numbers that changed

A dashboard shows where things stand, while a narrated animated chart shows what changed and why. Build one from two quarters of results, then consider making it a recurring self-updating video.

Questions to Consider

• Which explanation do you or your team give over and over, such as onboarding or customer education, that a self-updating video could replace?
• Who needs the information your team produces but will not read the doc, and what would a two-minute video version for them contain?
• Which recorded customer calls or research sessions hold insights your company's decision makers have never actually seen or heard?
• What product concept are you currently describing in text that would be clearer as a recorded walkthrough of a prototype?
• Which recurring weekly update, built from metrics, tickets or shipped features, could an agent render as a short video every Friday?

Bottom Line

AI agents can now watch, edit, animate and even record video from plain-English instructions, which makes video a practical format for everyday work. The opportunity is not to rival professional editors, but to use video where it was never worth the cost before, and to build pipelines so those videos stay current on their own.

Concepts to Explore

Agentic video understanding

This is a model's ability to work through long video files to find specific moments, rather than relying only on a transcript. It matters here because NLW calls watching the slowest part of editing, and Google claims its version is up to 88% more token efficient on long recordings.

Computer use agents

These are agents that operate a virtual computer and browser like a person. In this episode they matter because an agent can open a product page and record its own demo with an animated cursor, which removes the recording step for demos.

Tools & Products

Experiments to Try

Experiment: Meeting highlight video

• Hypothesis: for a missed meeting with a real debate, a two to three minute highlight video helps people understand the decision better than a written AI summary.
• Change: pick one recorded meeting with a meaningful debate and ask an AI to produce a short highlight video of the key exchanges.
• Test: send the video to some people who missed the meeting and the written summary to others.
• Measure: how many open or finish each format, and whether they can correctly explain the decision and the reasoning behind it.

Experiment: Weekly Friday video update

• Hypothesis: a short automated video update gets more attention from stakeholders than the usual written weekly update.
• Change: build a pipeline where an agent pulls the week's metrics or shipped features every Friday and renders a two-minute video.
• Test: run it for a few weeks alongside the written update.
• Measure: view rates compared with read rates, and how many follow-up questions or replies each format gets.

Notable Quotes

"The slowest part of editing was always the watching." (NLW)

"is anyone going to watch a 60 minute zoom that they weren't on? Absolutely not" (NLW)

"if a meeting isn't video worthy, maybe don't have that meeting, and maybe just do it via text in the first place" (NLW)

"a video is a video, a pipeline is an application that produces videos in some recurring way." (NLW)