Figma + GPT-6 Astra in Action: AI Autonomously Designs a Complete Flight Control System

In a collaborative demo between Figma and an AI model codenamed Astra — widely associated with GPT-6-level capabilities — the model was handed a formidable challenge: design a complete flight control system interface for a single-person short-range helicopter. The result surprised Figma's design team. The model ran continuously for nearly two hours, independently producing a full prototype — from design system and components to a multi-screen user journey. The demo reveals a threshold that AI is crossing in professional design workflows.

A "Moonshot" Design Brief

The setup was deceptively simple: the designer provided just a few hundred tokens of instructions, describing a flight control system for a single-person helicopter optimized for short commutes, and asked the model to define its own "hero flows" and user journeys.

Critically, the designer gave no step-by-step guidance — instead asking the model to "first figure out what a good prototype looks like, then go build it." This kind of open-ended instruction is precisely where current AI tools tend to fall apart: without clear constraints, models often produce fragmented, incoherent output.

Astra defied that expectation. The Figma designer described feeling "a level of confidence when using it that I've never experienced with any other model" — it had its own judgment about how to approach the task and didn't need extensive hand-holding. More importantly, it would "give itself time to understand a new task before actually starting work" — a plan-first, execute-second pattern widely credited for the higher quality of the final output.

"Hero flows" is a standard product design concept referring to a product's most critical user paths — the ones that best embody its core value proposition. This typically means the full journey a new user takes to complete a key task (such as planning a first flight or executing a takeoff). Unlike edge cases, hero flows represent the "main road" that design resources should polish first. Asking an AI to independently define this path means the model must not only understand the interface, but make a judgment call about what matters most in the product — a high-order decision requiring business understanding, not just visual generation.

Two Hours of Autonomous Work, One Complete Prototype

The model ran for approximately two hours and delivered not a single screen, but a full multi-screen product experience.

What impressed the Figma team most was the model's deep use of native Figma capabilities. It established a unified brand system, used components correctly to ensure consistency, and stitched work across multiple screens into a coherent user journey — showing how a user would experience the app as they moved between interfaces.

"I've probably never seen a model use that many Figma features and get everything right — the design system, the components, the full experience walkthrough." That achievement implies the model had to "think from fairly foundational principles," clarifying the prerequisites for correct construction before touching a single frame.

The gap between a vague few-hundred-token brief and a high-fidelity, interactive prototype capable of communicating a complete idea — that leap is exactly what the demo was designed to highlight.

Design Systems and Components in Figma are the core mechanisms for maintaining consistency across large products. A design system typically includes color variables, typography rules, and spacing tokens, while components are reusable UI units (buttons, inputs, nav bars) where changes to a master component automatically sync to all instances. Knowing to build this system proactively — rather than copy-pasting elements manually — is a key marker separating professional designers from beginners. An AI that spontaneously establishes this infrastructure and maintains it throughout suggests output that is approaching engineer-ready, collaborative quality, rather than rough sketches for reference only.

Mid-Task Steerability

Beyond autonomy, another quality repeatedly highlighted was steerability. The designer noted that even mid-execution, the model could receive new directional instructions and adjust course — something described as "particularly special."

For real design work, this ability to "redirect while running" is more valuable than a single perfect output. Design is never a linear process. Being able to intervene in real time and correct the AI's intent as it works means human designers remain in control — rather than passively waiting on a black-box result.

Figma's "Agentic Inside and Out" Strategy

The demo also reflects a broader product vision at Figma. The team lead framed the core question as "how to make Figma agentic from the inside out" — ensuring users have access to the best models within Figma's own product, while also making Figma callable no matter where a user's workflow originates.

This means Figma isn't betting on a single model. Instead, it's building an open model integration layer. For designers, AI capabilities are no longer bolt-on tools — they're deeply embedded in the design system, component library, and collaborative workflow.

"Agentic" comes from the AI concept of an Agent: a system that can autonomously form plans, break down subtasks, invoke tools, and iterate toward a goal without requiring explicit human instruction at each step. Unlike traditional "question-and-answer" LLMs, agentic systems proactively maintain context, track goal progress, and make their own choices at decision points. Figma's vision of being "agentic inside and out" means Figma serves both as an execution environment for external AI agents (accepting operational commands from outside AI systems) and as a platform with embedded agent capabilities — letting designers delegate entire workflow segments to AI for autonomous completion, rather than merely using AI to speed up individual steps.

What This Means for the Design Workflow

Looking ahead, the Figma designer's enthusiasm centers on how models like this can genuinely integrate into daily work: from initial spec, through messy creative exploration, to a refined and aligned point of view, all the way through prototyping and visual communication.

One value point stood out in particular — the ability to de-risk the most technically uncertain parts of a design in parallel with the design work itself. Traditionally, design and technical feasibility validation are sequential phases. AI makes it possible to run them simultaneously.

This parallelization is expected to "dramatically accelerate workflows, ultimately enabling teams to ship more product faster." Of course, an official partnership demo naturally skews toward showcasing best-case results. Whether the same smoothness holds up against the messy constraints, repeated revisions, and edge cases of real projects remains to be validated through independent testing. But one thing is clear: AI's role in professional creative tools is evolving — from autocomplete assistant to a collaborator capable of independently owning end-to-end tasks.