# Higgsfield built its creative agent because its own app needed a second screen > Source: > Published: 2026-09-16 13:08:11+00:00 # Higgsfield built its creative agent because its own app needed a second screen **Head of Product Axultan Alimkulov's team spent six weeks building a harness for models, visual memory, tools and human creative know-how.** By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff) ยท Published Primary source: [Higgsfield AI Newsroom](https://higgsfield.ai/blog/how-we-built-supercomputer) ## Why it matters Higgsfield is using visual memory, model routing and human-authored Skills to make its workflow layer harder to replace than any single image or video model. [Higgsfield's inside account](https://higgsfield.ai/blog/how-we-built-supercomputer?ref=runtimewire), published September 16th, describes how a four-person product and engineering group built the system underneath [Higgsfield Supercomputer](https://higgsfield.ai/supercomputer?ref=runtimewire), the creative agent Higgsfield rolled out in May. The origin was visible inside Higgsfield's own office. Creatives kept Claude Code, ChatGPT or Gemini open on one screen to research ideas, develop prompts and hold project context. Higgsfield occupied a second screen where they generated the images and video. [Alex Mashrabov](https://www.linkedin.com/in/amashrabov?ref=runtimewire), who previously ran generative AI work at Snap after selling his computer-vision startup AI Factory to Snap, had built Higgsfield around shortening creative production. His own staff still needed another vendor's chatbot to operate it. Higgsfield's account describes how Head of Product [Axultan Alimkulov](https://www.linkedin.com/in/axultan?ref=runtimewire), AI engineer Alen Sultanov, Head of Prompt Engineering Ruslan Syzdykov and software engineer Toktar Akhmetov turned that product gap into a six-week engineering sprint. Their answer was an orchestration harness that could plan work, choose among models, manage files and visual references, and recover when an underlying service failed. For Mashrabov, Supercomputer also advances a larger wager. Higgsfield is moving beyond selling access to image and video generation and trying to own the layer where a creator's instructions, assets and model choices become a finished campaign, film, website or app. That layer holds the workflow context and customer relationship even as the models underneath it change. ### The agent started as a product complaint Alimkulov arrived at Higgsfield after Higgsfield acquired his startup, Invicta AI. His background was in large language models and agentic systems, where chat had become the interface for directing software. Higgsfield's visual tools still exposed the mechanics: select a model, format a prompt, adjust controls and generate. That interface became harder to operate as visual models acquired more capabilities. A video request could require a character reference, product image, location, previous generation and model-specific syntax. Video also costs more and takes longer to regenerate than an image, increasing the penalty for a poorly translated request. "We understood this couldn't go on," Alimkulov said in Higgsfield's account. The product group had already tested narrower solutions. AI Director in Cinema Studio 3.5 knew about a user's assets and could help format prompts, though it could not run the generation itself. Higgsfield later added an agent to Canvas, its node-based workflow builder, so users could describe a sequence instead of assembling every step manually. Those tools remained tied to predefined portions of the workflow. Alimkulov argued that fixed pipelines also failed once users deviated from the expected sequence. "You cannot predict user behavior with a rigid flow," he said. ### Visual memory became the hard part General-purpose agents offered a template for research, code execution, browser operation and file handling. Creative production added a costly form of context: images and video that may need to remain available many steps later. Higgsfield illustrates the problem with a user saving a generated character as "heroine" and later asking for "heroine holding a cat." The request sounds simple. Higgsfield Supercomputer must resolve the word to the correct visual asset, retrieve the useful information and avoid repeatedly loading an entire project history into the language model. The difficulty compounds across multiple characters, locations, props, source clips and previous generations. Higgsfield built a visual reference system to carry those assets across tasks while limiting how much context must be passed through the model at once. The second technical problem was model-specific taste. A general language model can produce a plausible video prompt without knowing how Seedance should be instructed compared with Kling, or which sequence of actions tends to work for a particular commercial format. Syzdykov's prompt engineering group helped encode that operational knowledge into what Higgsfield calls Skills. A Skill can contain instructions for a task such as producing a user-generated product ad, a short-form explainer or a longer animated video. Creatives and prompt engineers can author the procedure while the agent adapts the inputs to a user's request. "A Skill is essentially know-how - knowledge a human has and an LLM doesn't," Alimkulov said. That design turns the work of Higgsfield's prompt engineers and creatives into reusable product infrastructure. It also gives Mashrabov a route to differentiate Higgsfield without depending entirely on exclusive access to a model. Competitors can offer many of the same underlying image and video systems. Reproducing the accumulated instructions, asset handling and production habits inside a workflow layer takes a different kind of work. ### The business sits above the models [Higgsfield Supercomputer currently routes tasks](https://higgsfield.ai/creator-hub/help-center/tools/how-do-i-use-supercomputer?ref=runtimewire) across generation systems and language models, with Higgsfield documentation listing providers including Anthropic, OpenAI, Google, xAI and DeepSeek. The workspace also includes persistent memory, file storage and connectors for services such as Slack, Telegram, TikTok, Gmail, Google Drive and Notion. Text interaction in the product's Free Mode does not consume credits, according to Higgsfield. Image, video and other generation steps do. A multi-step job can therefore use more credits than a single generation, and Higgsfield says the running cost is displayed as the task proceeds. That pricing decision reveals the commercial model underneath the interface. Higgsfield can make planning and conversation inexpensive while charging when the agent invokes the compute-heavy generation products at the center of Mashrabov's business. The timing matters. Higgsfield [raised a $400 million Series B](https://www.prnewswire.com/news-releases/higgsfield-raises-400-million-series-b-financing-at-5-4-billion-valuation-with-annualized-revenue-reaching-700-million-302852430.html?ref=runtimewire) at a $5.4 billion valuation on August 17th, led by DST Global. Goldman Sachs Alternatives, Tribe Capital, Smash Capital, Fifth Wall, Valor Capital, Intel Capital and other investors participated. Higgsfield said at the time that it had surpassed 30 million users and reached a $700 million annualized revenue run rate. Those are Higgsfield's figures. [Reuters reported in January](https://www.reuters.com/business/media-telecom/ai-video-startup-higgsfield-hits-13-billion-valuation-with-latest-funding-2026-01-15/?ref=runtimewire) that Higgsfield's earlier $200 million annualized figure was a projection rather than recognized revenue, a distinction that also applies when reading the later run-rate claim. The [August financing announcement](https://www.prnewswire.com/news-releases/higgsfield-raises-400-million-series-b-financing-at-5-4-billion-valuation-with-annualized-revenue-reaching-700-million-302852430.html?ref=runtimewire) attributed a 42-fold increase in users of Higgsfield's agentic products over the three months following Supercomputer's May rollout. Higgsfield did not provide the starting user count in that comparison, leaving the scale behind the multiple unclear. The announcement also said the company generated more than 20 million pieces of content per month. ### Higgsfield is turning the harness into its operating system A six-week first build did not finish the work. In [Higgsfield's account](https://higgsfield.ai/blog/how-we-built-supercomputer?ref=runtimewire), Akhmetov said the team tests Supercomputer and runs evaluations after major updates. The account also says feedback from Higgsfield's cinematographers and creatives, among the product's heaviest users, helps determine priorities. Higgsfield has since added Projects for shared context, plus workflows for apps and games. Supercomputer now runs inside Cinema Studio and Canvas. Higgsfield also exposes generation capabilities to outside agents through MCP, extending the same strategy beyond its own interface. The common thread is control over the job around generation. Mashrabov can buy access to new models, add them to Higgsfield and route tasks toward them. The harder asset is the production system that remembers what the creator is making, knows how each model behaves and turns a short instruction into a sequence of paid actions. That is the bet behind the grandiose Supercomputer name. Higgsfield is building the place where the models report for work.