Mission Control is the right direction. The proposal is thoughtful and unusually well tested, but the current film mistakes activity for progress and a cueable API for cueable agents. It would impress people who already understand Agent Manager more than people meeting it for the first time.
Do not lock this cut. Keep the visual direction and the real-capture principle; rebuild the centre of gravity around one visible outcome. The existing 66.8-second cut is a good technical proof and a weak release film.
The work is strongest when it measures and weakest when it infers. Its capture engineering is credible. Its story does not yet close the loop.
I checked the original Kimi post and full-resolution video, not only the proposal's clips.
| Claim | Finding | Assessment |
|---|---|---|
| Identity | The X post is titled Meet Kimi K3, published 16 July 2026. Its embedded count was 30,604 likes when checked. The source video is 1920×1080, 29.97 fps and 56.7467 seconds. | Verified |
| 26 cuts / 2.18 s | The arithmetic uses 26 as the number of shots: 56.7 ÷ 26 = 2.18. Twenty-six cuts would create 27 shots and a 2.10-second mean. More awkwardly, the proposal's own cut map draws 23 ticks. Automated detection returns 22 at a 0.24 threshold and 29 at 0.18 because animated transitions blur the definition. | Not defensible as written |
| Acceleration | The broad shape is right: longer setup, a dense run of roughly one-second swaps around the capability montage, then a hard brake. The final transition begins at about 50.72 seconds, leaving a six-second outro. | Directionally right |
| Ground explains calm | The sage desk is important, but it is not the whole reason. Consistent card placement, centred composition, repeated prompt labels, musical phrasing and the alternation of held shots with bursts do equal work. The middle is intentionally frantic. | Overstated causality |
| Music structure | ffmpeg's EBU R128 pass confirms a 6.7 LU loudness range. The quiet title interval and final decay are real editorial devices, not invented in the write-up. | Verified |
| “Confirmed” reference | The factual description of Kimi is strong; the inference about what the operator meant is still an inference. The page itself says “confirmed” in the masthead, “almost certainly” in §01, and “still unconfirmed” in §06. | Internally contradictory |
| Creality rival | An official Creality K3/KliTek campaign exists, but it says the K3 is coming in Q3 2026. The proposal gives no link for its exact 30 May “launch film” claim, so teaser/reveal is the safer description. | Plausible, weakly sourced |
Sources: original Kimi post · official Creality K3/KliTek campaign · proposal and first cut.
It is clearly the best of the three directions, but “best direction” does not yet mean “working film.”
| Direction | Critical read | Call |
|---|---|---|
| Paper Fleet | K3's charm comes from photographed paper, hands, shallow focus and material imperfection. A scan plus composited cards keeps the borrowed silhouette and loses the reason it worked. The proposal correctly identifies the homage risk; that risk is fatal, not merely a caveat. | Wrong direction |
| Mission Control | Honest, ownable and cheapest. The dark UI is not the problem. Scale is: at six panes, text is already unreadable and a 1 px frame around a tiny icon is peripheral motion. At 47 rows it would be a screensaver. The direction needs visible cause and consequence, not more activity. | Keep, but refocus |
| The Handoff | The operations graph is the most unique image in the proposal because it describes something the product knows. Twenty seconds of it would sell a diagram instead of an app. Three seconds as connective tissue is sensible, provided the edges are genuinely sourced and legible. | Supporting device only |
Mute the captions and ask what the five-pane shot says. The answer is “several terminals are active.” It does not say they share a task, that one hired the others, that review has begun, or that anything is closer to done. Cursor's subagent film works because a count climbs and files accumulate. The proposal notices that lesson, then does not use it.
A fleet is impressive when work visibly moves between members and converges. Many animated borders are only evidence that many processes exist.
The 66.8-second first cut is valuable because it lets the plan be judged against real pixels rather than prose.
| Beat | What a new viewer can actually read | Verdict |
|---|---|---|
| 0:00–0:12 launcher / group | A dark, already-busy admin UI zooms out. With the pointer deliberately hidden, selections, drag-and-drop and grouping appear to happen by themselves. K3 earns its slow opening with a human hand and an image forming; tiny CLI marks do not have equivalent pull. | Too long and too procedural |
| 0:16–0:21 filament graph | A generic network diagram with labels too small to parse at feed size. It says “connected things,” not yet “agents handed work to one another.” | Useful only as a bridge |
| 0:21–0:38 fan-out / fleet | This should be the film's proof. In the cut, panes reflow and the number of sessions changes, but the hiring event is not visually dominant. The proposal asks three agent creations, streaming work, a wait completing and a review starting to read within fourteen seconds. That is several ideas, not one readable beat. | Core story is compressed |
| 0:38–0:46 files / research | A concrete file opens and contains a real, structured comparison. This is the strongest product evidence in the cut. But it is framed as “what they wrote,” not a clean answer to the benchmark question. | Earns its time; needs payoff |
| 0:46–0:51 phone | The responsive layout is genuine. A tall phone containing the same dense conversation is still tiny text inside a frame, and without a hand or visible pointer the claimed reply is easy to miss. | Proof, not a human moment yet |
| 0:51–0:56 three settings | Usage, Skills and General at roughly 1.3 seconds each are icons of features, not demonstrations. With no voice-over there is not enough time to read both the caption and any value in the UI. | Feature salad; cut from main film |
| 0:56–1:01 remote | The current cut shows only a concept graph because no remote agent was connected. That is exactly the line between a product claim and product footage. | Do not ship until real |
| 1:01–1:07 end card | The line lands, the hold is welcome, and it makes the production story part of the product story. | Keep |
The page describes the same plan as 62 seconds, then 66 seconds, while the embedded cut is 66.8. It also alternates between a 4.8 and 5.1-second mean. These are editorially small errors, but they matter in a document arguing from measured pace.
The technical plan proves a reliable recorder, not a frame-accurate fleet.
This can be a real, staged demonstration assembled from real event-driven takes. It cannot be a frame-accurate live choreography unless the director waits on DOM/event barriers and lets the edit—not the agents—decide when each beat lands.
The 60 fps target also lacks a product reason. Kimi itself is 29.97 fps. For a mostly static UI, 30 fps reduces capture and encoding pressure; 60 fps should survive only if a measured side-by-side shows that the tiny border motion materially improves.
The missing release-video beat is not another feature. It is the result of the story already underway.
The film asks “what is the fastest way to resize 100 JPEGs?”, shows research, hiring and activity, then changes the subject. It never gives a crisp answer, a measured win, a reviewed change, a passing test, or a merged artifact.
That omission creates the larger positioning problem: a newcomer can reasonably conclude that Agent Manager is a polished terminal tiler with animated presence. The product's stronger claim is that work can be delegated, observed, handed back, reviewed and resumed across sessions and machines. The current film shows delegation and observation. It does not show convergence.
| Missing proof | Why it matters |
|---|---|
| Outcome | The benchmark needs one legible result connected to the opening prompt. This is the difference between a fleet that burns tokens and a fleet that completes work. |
| Closed handoff | We see peers appear, but not a visible review finding being returned and resolved. That loop is the “manager” in Agent Manager. |
| Continuity | The phone implies portability, but not that the same durable session survives leaving the tab, a sleep or a restart. Without one continuity cue, the product can be mistaken for disposable multiplexed terminals. |
| Readable causality | Raw POST lines are implementation detail. A newcomer needs to see who asked whom, what arrived, and what changed because of it. |
Direction 2 is not wrong; the current evidence hierarchy is. Make the real result the hero, make fan-out a cause-and-effect step on the way there, and demote ambient status motion, settings and unproven remote claims. Until then, the first cut is a capable making-of reel for insiders—not the release film.
What I checked, and what remains judgment.
Inspected the full proposal and embedded 66.8-second first cut; downloaded the original 1920×1080 Kimi source from the cited X post; used ffprobe for stream metadata, ffmpeg scene scores at multiple thresholds, and EBU R128 analysis; inspected half-second contact sheets for both films; checked the current Agent Manager web source for tree and Overview polling cadence. I did not stage sessions, mutate the dev Space, or touch app code.
High confidence: metadata, arithmetic inconsistency, timeline tick count, loudness range, polling latency, capture-vs-behaviour distinction, and what is visible in the current cut. Judgment: whether a new viewer finds a beat impressive and whether the benchmark task justifies a fleet. Not established: which K3 reference the operator intended, and the exact cut count under a single agreed definition of animated transition versus edit.