Krea Bots: Nine Agents, One Shared Memory, and a Waitlist

On November 12, Krea opened a beta for Krea Bots. The pitch on the page is short: an AI teammate with a face, a craft and a memory. There are nine of them. One plays creative director and is called North. The rest cover storyboard, images, video editing, title animation, sound, social, 3D, and brand guidelines.
You work with them in a conversation. You mention a bot with an @ to hand it a task, it does its part, and it passes the next step to whoever owns that craft. Access is through a waitlist, and Krea is giving teams it admits $1,000 in credits. There is no published price and no release date.
Krea also says all of its recent content was made with these bots, including the video in the announcement. That claim cannot be checked from outside. Beta launches use that kind of claim to stand in for a benchmark, so start there.
The memory is the actual feature
Generation is not new. Krea has had generators for three years. What the bots add is a place to put decisions that currently get retyped every single time.
Before a bot starts, it reads three things: the lessons it has learned from your feedback on its craft, the notes from the current project, and the workspace house style. Krea's own example of a house rule is "logo bottom right". Lessons a bot proposes wait for your approval before they become permanent.

Anyone who has produced visuals for a brand at volume recognizes the tax being described here. The brief is not the expensive part. The expensive part is the twentieth image, where you have to say again that the logo goes bottom right, that the light is late afternoon, that the palette stops at three colors, and that the type is never condensed. That information lives in a PDF nobody reads and in the head of one person who is on holiday.
An agent that keeps those rules answers a real problem, and it answers it in the least glamorous way possible: by remembering a sentence you said once.
The three-layer split is better than a single blob of context. Feedback lessons are about taste and repeat mistakes. Project notes are about this campaign. House style is about the company. Mixing them is how you end up with a bot that applies a one-off note from a client revision to every job for the next six months.
The tools list deserves attention
Each bot also carries tools. Krea's generators, obviously. The page and AlphaSignal's coverage also list Slack, Meta Ads, Figma, and a browser. Each bot looks at every image it makes before it replies.
The browser is the item to notice. A creative agent that can look things up can check whether the typeface in a reference is actually the typeface the brand uses, or whether a competitor already ran this campaign. A creative agent that cannot look anything up will confidently produce a moodboard of things that do not exist.
Meta Ads is the second item to notice. It suggests the intended loop closes on performance rather than on approval. Generate the variations, ship them, and let the bot learn which crop held attention. That is a different product from a design tool, and it moves Krea toward the media buying side of the house.
The numbers Krea put on the page
The example week goes like this: three creative directions, twenty product photos, an animated turntable, an ambient sound bed, a fifteen-second spot, and posts cropped for each network. The total shown is 43 hours of bot work for seven minutes of yours.
Treat that as an illustration. Krea labels it as an example, and there is no published method behind the 43 hours. The seven minutes is the more interesting half anyway, because it implies the human contribution is reduced to reading and approving. Whether that is enough depends entirely on what those seven minutes contain.
If the seven minutes are "pick one of three directions" and "approve", then the work that disappeared was execution, and the work that remained was judgement. That is a good trade for most teams. If the seven minutes are "fix the crop the bot got wrong eight times", the number is marketing.
The control problem nobody has solved
Nine agents handing work to each other creates a specific failure that a single model does not have. When the output is wrong, you need to know which agent was wrong and which agent inherited the mistake.
Krea's answer is approval. A bot shows what it is about to do, and you keep the decisions. That is honest, and it is also the cheapest possible answer. Approval screens are insurance against runaway generation, and they work best exactly where they are least needed: short, well-specified tasks. They work worst on the long chains, where a wrong assumption entered at step two and only becomes visible at step seven.
The deeper issue is that a house style written in ten lines cannot cover the cases where two rules conflict. The logo goes bottom right, and this poster has no bottom-right space. A human art director resolves that in two seconds using taste and context. An agent with three memories either follows the rule and produces something ugly or drops the rule and produces something off-brand.
That is where the seven minutes goes, if the product is good. The approving is the easy half. The exceptions are the work.
Why this lands now
Krea's own November is busy. It added GPT Image 2.6 on November 6 with faster generation, single-reference editing, and image-to-GIF. It relaunched Krea Video on November 5 with camera controls, more than a hundred presets, and sound generation. It shipped a Premiere plugin in late October. Multiplayer arrived in Krea Nodes on October 30, team features on November 9. Bots on November 12.
Read as a sequence, that is a company building the surface where a team works before it builds the thing that works on the surface. Multiplayer puts several people in one project. Team features put a team in one account. Bots put a colleague in the conversation.
The direction is the same one Adobe, Google, and DomoAI are moving in, from different starting points. Adobe is putting generation inside an editing timeline that already holds the context. Google is putting generation inside a set of artist-designed tools in Flow. DomoAI shipped an agent on November 13 that takes one sentence to a finished vertical cut and stops to ask for approval before each stage.
All four are conceding the same point: the model is no longer the differentiator, and the surrounding decisions are.
What I would do before joining a waitlist
Write the house style down. Ten lines: logo placement, palette limits, the kind of light you use, the words you refuse, the type you never set, what a crop may cut. Do it whether or not you get into the beta, because that document is what any agent will read, and it is also what a new freelancer will read on their first day.
Then pick one campaign, not your whole workflow. The failure mode for agent products is a team that hands over everything in week one, gets one bad batch, and concludes the category is hype. A single campaign gives you a comparison you can actually make: same brief, same team, bots on one half.
And keep the approvals on, because a chain of nine is nine places to lose the thread. The one thing you cannot delegate is knowing which frame ships.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.