Le nouveau flux de travail d’IA agentique semble trop puissant

I expected something like this to come sooner or later, especially the agentic side of it because that is where everything in AI seems to be heading right now. It is no longer just about generating one answer, one image, one video, or one app screen. More and more, these systems are starting to act like agents that can understand a goal, follow context, use different tools, move through steps, and turn a rough idea into something more complete. And that is why this new Abacus release caught my attention. At first, it looks like a design update, but it is actually broader than that. Abacus is showing two connected things here. The new design vertical inside Abacus AI agent and a Gentic media generation inside Abacus Studio. The design vertical is focused on things like turning rough sketches into app screens, creating user journeys, building wireframes, designing mobile apps, generating enterprise dashboards, and exploring brand identity direction. and studio goes into the media side. Product videos, horror, web comic clips, animated characters, motion transfer, cinematic landscapes, subject consistency, image editing, upscaling, and finished campaign style assets. And the more interesting part is that they are trying to connect the creative process itself. Instead of AI stopping after one output, it can move through multiple creative steps and keep the original direction alive across the workflow. And once you look at the demos, that agentic angle becomes pretty obvious. Abacus AI agent now has a dedicated design vertical. And the interesting part is not just that it can generate screens. A lot of tools can do that. The interesting part is that Abacus is trying to make the agent reason about design before it generates anything. It looks at the product, users, tone, goal, and visual language, then builds around that instead of just throwing a stylesheet onto a layout. And you can see that right away in the first demo. They take a rough handdrawn sketch of a travel app with four screens. A discover page, an itinerary view, a map, and a profile. It has annotations, arrows, location names, pin colors, and messy labels. Then they upload it into Abacus AI agent and type something simple like convert this into design. The agent reads the sketch, picks up the arrows, labels, locations, pin colors, and implied navigation, then generates a full Python script, opens a canvas, and turns the rough sketch into highfidelity screens. That matters because many product ideas start as rough notes and half-drawn screens, not polished Figma files. If AI can translate that into something clear enough to build from, that is a real workflow improvement. The second demo goes into user journeys for a credit card application platform. The prompt asks the agent to create wireframes that visualize the onboarding flow for a credit card application. A basic AI tool would probably give you five or six screens. Sign up, personal details, documents, approval, maybe a dashboard. Abacus creates 30 screens, 15 for web and 15 for mobile. The agent also breaks the experience into different flows. There is the happy path where everything goes smoothly. There is a pre-qualification path for users who might not qualify. There is a save and resume flow for people who start the application and come back later. And there is a full error state structure for when things go wrong. Users mistype things, abandon forms, fail checks, get interrupted or need reassurance. So in this case, the AI is not just drawing screens. It is mapping the experience. Then there is the luxury sports club app demo. The prompt is basically, "Help me design a mobile app to manage a luxury sports club." The agent asks who the users are, whether it is for members, staff, or both, and what the aesthetic should feel like. Then it builds a design system, deep navy, antique gold, champagne accents, Georgia serif paired with interan serif. It even writes dos and don'ts for the visual language, so the design has rules instead of random colors. From there, it creates seven screens across two experiences. For members, there is a splash screen, a home dashboard greeting the user by name with good morning, James Harrington, facility booking with time slots and add-ons, class registration with featured workouts, and a profile with a digital membership card and barcode. For staff, there is an admin dashboard with four KPI cards, a facility occupancy chart, severity coded alerts, a live activity feed, and a member management screen with search, tier filtering, and one-click approve or decline actions for pending applications. So, from one prompt, it builds both the member side and staff side while keeping the same luxury feeling across everything. The hospital operations demo is interesting because healthcare design is easy to get wrong. The prompt asks for a healthcare operations app that feels calm, trustworthy, human, and polished while avoiding generic dashboard aesthetics. The agents design system describes the product like a trusted colleague, clear, composed, and never alarmist. It even describes the emotional tone as quiet confidence. The app is called Meridian, and it includes eight screens. There is a main operations hub that greets Dr. Lynn by name and shows what needs attention. There are patient flowcharts, a color-coded bed management grid, a staffing risk heat map with fatigue tracking, and an AI recommendation screen with confidence scores, impact metrics, and accept or dismiss actions. The pallet also makes sense, muted, teal, and green with red reserved only for real emergencies. Another demo shows a normal design workflow. Low-fidelity wireframes into highfidelity screens. First, the user asks for mobile app wireframes. The agent creates 10 grayscale screens for a book app called Bookshelf. Welcome, signup, home feed, discover, book details, shelving, library, reading progress, reviews, and profile. It also creates a navigation flow diagram. Then the user says, "Now convert this into highfidelity design." The wireframes turn into a warm terracotta and cream app. It sources 14 real book covers from Open Library, including Dune, The Midnight Library, Atomic Habits, and Project Hail Mary. Then it adds progress bars, star ratings, color-coded shelf icons, success toasts, and bottom sheet drag handles. That is a proper workflow. Structure first, polish second. The last design demo is brand reinvention. The prompt is reinvent the brand identity for abacus.ai. Before designing anything, the agent opens a browser, visits the abacus website, scrolls through it, extracts the logo, reads the homepage sections, and even runs JavaScript in the browser console to pull exact hex codes for the brand colors. Then it creates a brand research document with colors, typography, content structure, design tokens, and overall identity. Only after that does it create four landing page directions. One is vibrant and new age with dark purple gradients, glowing CTAs, bold stats, and high contrast. One is minimalistic muted orange with clean white space, muted orange accents, monospace dividers, and offset shadows. Another uses soft muted colors, lavender, rainbow pastels, aricolor gradient headline, and feature cards in different color families. The last one is clean black and white, sharp, crisp, and hyper minimalist. Same brand, same content, same logo, four different directions. Now, the other part of this release is Abacus Studio and Agentic Media Generation. We all know that AI media workflows can get messy fast because making one polished 30-second asset can still mean jumping between separate tools for images, video, voice, editing, animation, and upscaling. Abacus Studio is trying to put that workflow into one environment. You describe the outcome and it helps move from idea to image, image to edit, edit to video, video to upscale, and concept to finished asset. It uses video models like Cedence, Cling, and Veo 3. Image and editing models like Flux.2 Pro and GPT image 2, plus workflows for upscaling and enhancement. The first media example is an AI product review video. You describe the product, choose a tone or style, and the platform generates a complete product video with visuals, motion, and voice over. You can include brand elements like logos, colors, and messaging so the final clip matches the product identity. For marketing teams, that is useful for product showcases, ads, and social media content where speed and iteration matter. The second example is a horror web comic video. It starts with a dark, grainy comic panel idea. an abandoned hallway, a red hooded figure, glowing red eyes, peeling walls, heavy shadows, and a claustrophobic mood. Abaca Studio turns that into a 47.9 second video at 2560x440 resolution. It adds a slow camera push, character movement, narration boxes, comic panel transitions, grain, static effects, eerie sound design, dramatic pacing, and a jump scare style ending. The story builds around the line. They said the hallway was abandoned for years. No one went in. No one came out. Then the redeyed figure gets closer. The text becomes shorter and colder. Footsteps and breathing build tension. Static hits. And it ends with the hallway. It never ends. So it is assembling a short form story with mood, motion, pacing, text, and sound. The third demo is motion transfer. It starts with an anime style character with long blue and orange hair, expressive eyes, a colorful oversized outfit, a patchwork hoodie, multicolored jeans, sneakers, and a pastel background. Then the user uploads a liveaction video of a real dancer doing arm waves, body rolls, bounces, dabs, flexes, and expressive movements. The instruction is simple. Transfer the uploaded video motion to the character image. The result is a 35.1 second 2560x1440 video where the character performs the dancer's movements while keeping its identity stable. That means the workflow handles character generation, video ingestion, motion understanding, pose transfer, character preserving animation, and high resolution output. For brands, mascots can move, campaign characters can perform, animated presenters can host, and one human performance can become multiple animated variations. The fourth example is a cinematic nature scene. It starts with a hyperrealistic image prompt. Iceland waterfalls or Norwegian fjords, golden hour lighting, professional photography, wide-angle composition, atmospheric depth, and a premium documentary feeling. The image is generated with Flux.2 2 Pro. Then the user asks for stronger god rays, richer golden hour light, drifting mist and fog, more powerful water flow, layered mountains, dramatic clouds, natural but more vibrant colors, subtle birds, and a BBC Earthstyle documentary look. After that, the still image becomes a 35.4 4 secondond video at 25560x440 with smooth drone-like camera movement, flowing waterfalls, drifting mist, subtle cloud motion, birds in the distance, shifting sunlight, ambient nature sound, and a wide vista ending. The workflow also includes a two times upscale from 1,280x720 to 2560x440. 60 frames per second enhancement and Topaz AI upscaling. That kind of output could work for website loops, trade show screens, product backdrops, luxury campaigns, keynote openers, investor presentations, and corporate videos. The fifth example is the peacock demo. And this one is mostly about consistency. It starts with a hyperrealistic Indian peacock generated like a professional DSLR wildlife photo with true-to-life color, iridescent feather detail, clear feather geometry, and detailed oselli, which are the eye spots in the feathers. This uses flux 2 pro. Then the same peacock is moved to a grand castle porch with stone flooring, arches, and marble columns while preserving the bird's identity, proportions, feather arrangement, lighting, shadows, and ground contact. That edit uses GPT image 2. Then the same peacock becomes a 34.2 second cinematic video walking across the porch. The prompt asks for strict temporal consistency, including feather count, shape, structure, proportions, realistic gate, head bobbing, tail sway, surface contact, and a stable environment. That is the harder part of AI media, keeping the same subject consistent across edits, locations, and motion. For products, mascots, models, campaign worlds, and brand identities, that matters a lot. So, the bigger point here is that Abacus is clearly moving away from one-off generation and toward full creative workflows. One nice UI screen or one cool image is not enough anymore. The value is in keeping the intent alive across the whole process. Creative judgment still matters because someone has to know which direction fits the product, but the starting point gets much stronger. Anyway, let me know what you think about Abacus moving into design and agentic media generation. Thanks for watching and I'll catch you in the next one.