Skip to content

Inside Higgsfield #2: How We Built Supercomputer

2026-09-16 · BTS / making-of / vlog · Higgsfield, Axultan · original

Topic: Higgsfield platform & ecosystem (also: Prompting & AI assistants, Cost, credits & plans, Video generation & shot direction) · mode: Case study · tags: series, guest or collaboration, vendor claims

A behind-the-scenes post from 16 September 2026 on how Higgsfield built Supercomputer, its agent workspace for images, video, audio, code and files, in about six weeks. It traces the path from fixed prompt pipelines, which Axultan says broke down as models improved, to an agent that can call a character by name. The hardest problem was context, which led to a visual memory that finds the right saved character without reloading every image, and to Skills that carry the know-how of daily users, such as how to prompt Seedance versus Kling.

What is said

  • One system underneath Higgsfield

What is said

★ — how much it matters to the source.

  • ★★★ Higgsfield describes Supercomputer as an agentic workspace that researches, writes code, works with files and creates images, video and audio in one place, on a harness built for visual context, many models and Skills; its first version took about six weeks to build. · Supercomputer
  • ★★ Higgsfield's creatives commonly had Claude Code, ChatGPT or Gemini open on one screen to research, work through ideas, improve prompts and track a project, and Higgsfield on the other to generate. · Claude Code, ChatGPT, Gemini
  • ★★ Axultan, Higgsfield's Head of Product, who came from LLMs and agentic systems where chat was the natural interface, found it strange that image and video generation still meant choosing a model, writing a prompt, adjusting controls and generating.
  • ★★ As image and video models became capable of full campaigns and films, Higgsfield's own creative team spent more time outside Higgsfield, because prompting properly and carrying a project's context by hand was getting difficult. · Higgsfield AI
  • ★ Axultan says Higgsfield understood that working with one system to think through the work and another to make it could not go on.
  • ★★ For Ruslan Syzdykov, Higgsfield's Head of Prompt Engineering, the challenge in building Supercomputer was translating ordinary language into the structured instructions Higgsfield's models need. · Supercomputer
  • ★★ Before Supercomputer, Higgsfield and its prompt engineers built quiet pipelines that improved requests before they reached image and video models; they helped but were rigid, always running the same fixed steps in order. · Supercomputer
  • ★★★ Axultan says user behaviour cannot be predicted with a rigid flow, and the better the models became, the more obvious that limit of fixed prompt pipelines was. · agentic AI
  • ★★★ Axultan calls Seedance 2.0 the major shift: the earlier workflow animated a keyframe used as start or end frame, while Seedance 2.0's reference capability lets you tag a character with @ in the prompt instead of describing it again and again. · Seedance 2.0, keyframe
  • ★★ In Axultan's view, Seedance 2.0's @ references led directly to Elements in Higgsfield: reusable characters, locations and props called back into new generations. · Elements, Seedance 2.0
  • ★★ Higgsfield says that as models did more, creators had more to manage: video was slower and more expensive to iterate than images, Elements had to be referenced in the right format, and the prompt had to carry assets, references and how to use them. · Elements
  • ★★ Higgsfield's first attempt at an agent was AI Director in Cinema Studio 3.5: it knew the user's assets and Elements and how to format the prompt, but could not generate anything itself. · Elements
  • ★★ Canvas, a workflow builder, got an agent next: Higgsfield says describing what should happen in ordinary language can be much easier than assembling the same workflow by hand with nodes. · Canvas
  • ★★ General-purpose agents such as OpenClaw and Hermes, which Axultan calls superagents, could research, work with files, operate browsers, write code, coordinate other agents and carry out long sequences of work from a simple request. · OpenClaw, Hermes Agent, agentic AI
  • ★★★ Higgsfield says a visual project accumulates characters, locations, product images, reference footage and earlier generations, and passing all of it back into an LLM does not scale: images take a lot of context, video far more.
  • ★★ When improving AI Director or the Canvas agent stopped looking like enough, Higgsfield started building its own harness, the layer that coordinates the models, tools, context and steps behind the work. · Canvas
  • ★★ Higgsfield's harness was built to connect an agent to its generation, carry visual context across image and video, work across models and stay reliable when one part fails, drawing on self-improvement ideas from Hermes and enterprise controls from NVIDIA. · Hermes Agent
  • ★★ Higgsfield says its prompt engineering connects the agent harness to what its models can do, so someone asking for a cinematic video need not decide the prompt's structure, which capability to use or how to turn the idea into model instructions.
  • ★★ Alen Sultanov, an AI engineer at Higgsfield, describes Supercomputer's goal for users: arrive with a basic idea and still get a strong result without first learning prompt engineering. · Supercomputer
  • ★★ For Toktar Akhmetov, a Higgsfield software engineer, reliability was part of the problem: models and infrastructure can fail, and Supercomputer has to recover so a user's generations continue instead of collapsing with one failed dependency. · Supercomputer
  • ★★★ Higgsfield says the hardest technical problem was context: after a character is saved as "heroine", "heroine holding a cat" must reach the right visual asset without reloading the whole history of images and video into the LLM, so visual memory is core to Supercomputer. · Supercomputer
  • ★★★ Higgsfield says a general-purpose LLM may know how to write a prompt but not how Seedance should be prompted compared with Kling, or what a creative workflow needs; that knowledge sits with daily users, so Higgsfield built Skills to carry it. · Seedance, Kling

  • ★★★ Axultan defines a Skill as know-how: knowledge a human has and an LLM does not.

    “A Skill is essentially know-how — knowledge a human has and an LLM doesn’t”

  • ★★ Skills turn the know-how of Higgsfield's creatives and prompt engineers into agent instructions; for Marketing Studio and a YouTube feature in the works, creatives and prompt engineers write the instructions themselves: what happens first, next, and how it is generated. · Marketing Studio, YouTube

  • ★★ Axultan says a Supercomputer workflow needs no buttons for each choice, such as red or blue light: you say it in ordinary words and it understands. · Supercomputer
  • ★★ Higgsfield changed Supercomputer so that Free Mode text interaction costs zero credits, while image, video and other generation steps still use credits. · Supercomputer, credit
  • ★ Higgsfield says its core value is creative generation, and the conversation around it should make it easier to research, plan, write and direct before a generation happens. · Higgsfield AI
  • ★★ Higgsfield says Supercomputer now works across writing and research, code, websites and apps, PDFs and presentations, images, video, voiceovers, music, marketing and connected services, starting from a goal or from a file, image, video, document or URL. · Supercomputer
  • ★★★ Toktar Akhmetov says Supercomputer now runs inside Cinema Studio and in Canvas. · Supercomputer, Cinema Studio, Canvas
  • ★ Higgsfield says MCP gives external agents another way to work with its generation capabilities. · Higgsfield MCP
  • ★★ The Supercomputer team tests the system and runs evals after major updates, which has improved image handling and video analysis; feedback from Higgsfield's own cinematographers and creatives, some of its heaviest users, sets what comes next. · Supercomputer
  • ★★ Recent Supercomputer releases added Projects to keep shared context across conversations, Apps and Games, and new models including GPT-6 Astra; after users asked for game creation, the team built the infrastructure and shipped Games in two days. · GPT-6 Astra, Supercomputer
  • ★ Higgsfield's cinematographers use Supercomputer and Cinema Studio to create Higgsfield Originals. · Supercomputer, Cinema Studio, Higgsfield Originals
  • ★★ Higgsfield says the next Inside Higgsfield story will go behind the scenes of how the Higgsfield Originals films and series are made. · Higgsfield Originals

Threads in this video

Where the stories that moved here went next

Mentioned

Central: agent harness · Skill · Supercomputer

Also: People: Alen Sultanov · Axultan · Ruslan · Toktar Akhmetov

Projects: Higgsfield Originals

Organizations: NVIDIA · YouTube

Concepts: agentic AI · keyframe

Products: AI Director · Canvas · ChatGPT · Cinema Studio · Cinema Studio 3.5 · Claude Code · Elements · Gemini · GPT-6 Astra · Hermes Agent · Higgsfield Apps · Higgsfield Gaming · Higgsfield MCP · Kling · Marketing Studio · OpenClaw · Seedance · Seedance 2.0 · Supercomputer Projects

Also centred on Skill: AI Creation: First Steps: Product Images and Ads

Replies to

← Previous: What is the Higgsfield API, and how do I get started? · Next → How to Turn Sports Product Photos into AI Video Ads in 2026