Higgsfield - Building Agents That Provision Their Own Compute | DevDay 2026
Alex Mashrabov, CEO of Higgsfield, explains how agentic systems choose among models and provision their own GPU infrastructure to deliver consistent, multi-step creative media. He covers the trade-offs between speed, controllability, and dedicated deployment, and shares the 10-hour AI-generated live stream as an example of what becomes possible for small teams.
Overview
Alex Mashrabov, co-founder and CEO of Higgsfield, walks through how his team built agentic systems that select models and provision their own GPU infrastructure, with the goal of making continuous, multi-step media creation reliable enough for production. He frames the work as a response to a gap he saw repeatedly: generating a strong concept image is easy, but delivering a consistent experience around it is hard.
Product Context and Motivation
Higgsfield positions itself as a multimedia AI platform where creative marketers build end-to-end advertising campaigns using models such as OpenAI’s Sora (transcribed as “Astra” in the transcript) alongside image and video models hosted by Higgsfield. Mashrabov stresses that these AI-generated ads are meant to drive measurable revenue for businesses, not just produce visual assets.
He also points to unmet demand for new forms of interactive media that were not practical before. The Higgsfield API was launched to let developers ship those experiences, so that a team of one or two people can build an interactive show that would have been hard to imagine a few years ago.
Why the Agent Was Needed
Mashrabov describes a video editing skill built at Supercomputer that can produce a consistently coherent 10-minute video. That capability required significant spend, reportedly over $30,000, to run as a production-grade experience. Making such a workflow repeatable forces many decisions: which model to use for which step, what quality bar an output must meet, and how to know when a result is acceptable.
This is where he argues agentic systems must be strong at self-evaluation. The agent needs explicit, machine-checkable criteria for acceptable output so it can decide whether to re-render, switch models, or move on.
Creativity, Determinism, and Infrastructure
A recurring theme is that creativity begins with an idea but depends on turning that idea into a precise execution plan with clearly defined success criteria. Mashrabov says deterministic workflows are especially valuable here. These workflows may include advanced rendering with fine-tuned video models, so the team built an agent capable of launching its own GPU infrastructure. He credits this capability with making execution consistent across use cases.
Computer Use and Sub-Agents
He sees the new generation of models as the first to unlock real computer use. Traditional creative pipelines have required teams of 15 or more people fluent in 10 to 12 tools, and the cost of coordinating those tools acts as a “production tax” between an idea and its execution. In his view, a model that can drive those tools on the creator’s behalf, through sub-agents, lets smaller teams compete with the budgets of established studios, particularly in gaming, where visual quality expectations and budgets have grown sharply.
Trade-offs: Speed, Control, and Deployment
The main trade-offs he describes involve speed and controllability. Real-time experiences demand low latency, and he notes that recent ultra-fast model announcements are making that more accessible. For larger-scale or enterprise use cases, controllability becomes central. The team placed guardrails around the agent and deployed it inside a secure perimeter using dedicated GPU infrastructure. He says this approach accelerated enterprise go-to-market efforts, and that the same capability is now being exposed through the Higgsfield API.
Case Study: The First AI-Generated Live Stream
As a concrete example, Mashrabov describes a live stream built with a popular streamer, Neon, and a Higgsfield spokesperson. The two walked through New York City interacting with real people, while the model reconstructed the environment in a way that felt realistic. The stream ran for more than 10 hours. He presents it as evidence that a new media format, once impossible, is now within reach of a team of two or three people.
Advice for Developers
Mashrabov’s guidance is to build with the latest models and anticipate where capabilities are heading, echoing the idea of skating to where the puck is going. He is enthusiastic about the democratization of interactive media creation, which he attributes to models that can coordinate many tools. As an example, Higgsfield open-sourced an AI video editing skill used to produce explainer videos, and he notes that developers are now applying the same approach to tutorials for their own products. He closes with the point that Higgsfield was built in about 18 months and that this is the best time to start building.