Key Takeaways

  • Tangible spatial tools beat abstract benchmarks: users care more about generating Blender meshes and remodeling rooms than solving the Navier-Stokes existence problem.
  • Computer use models like Astra excel at friction-heavy administrative tasks, such as finding hidden control panel settings or queueing up parking ticket payments.
  • Physical robotics experiments show visual models adapting in real time, like an Astra-guided robotic arm painting the Golden Gate Bridge with iterative self-correction.
  • The shift in software interaction moves away from manual point-and-click interfaces toward delegating operating system controls directly to background agents.

Spatial Creation Beats Abstract Benchmarks

Tech circles spent years obsessing over whether artificial intelligence could crack Millennium Prize math problems like Navier-Stokes. Yet when real users get their hands on new generative tools, their attention immediately shifts somewhere else entirely: spatial creation and tangible outputs.

As John Coogan observed, user enthusiasm spiked when people started exporting 3D models directly into Blender, building web games, and redesigning their living rooms over a single weekend. “The weekend was very much Blender and people showing 3D renders and remodeling their houses and building small games and web games and that feels much more like a breakthrough,” Coogan noted. “People are having fun. They're vibe coding stuff but they're also doing 3D modeling which is like a new skill.”

Founders often mistake technical difficulty for product appeal. Proving an abstract theorem wins academic prestige, but letting a non-technical user generate an interactive 3D asset in five seconds wins distribution. Spatial output makes model capability obvious to anyone with a screen.

Automating the Boring Corners of the Operating System

Beyond 3D generation, agentic computer use models like Astra are attacking low-level desktop friction. Instead of clicking through complex settings panels, hunting down terminal commands, or configuring developer environments, users can offload system interactions directly to an agent.

Coogan pointed out that Astra is “particularly faster at computer use for like really really mundane things. Also just interesting to just have a very different interface to your computer like configurations all the things where you just have to dig through a bunch of settings maybe set something on the command line maybe find a control panel you can just ask codex or any AI agent to go and do it.”

The real product unlock is handling asynchronous real-world bureaucracy. Jordi Hays highlighted a common consumer headache: “Being able to go into chat and say, hey, once this ticket is live in the system, please pay it.” The value of computer use agents is not writing prose; it is closing the gap on tedious administrative loops that humans delay for weeks.

From Software Agents to Physical Control

The jump from operating system automation to physical execution is happening faster than expected. In one experiment highlighted by Coogan, researchers handed Astra a paintbrush, a camera, and a robotic arm, tasking it with painting the Golden Gate Bridge in the physical world. The system interpreted visual feedback, adjusted its mechanical movements, and steadily improved across repeated attempts.

As software models learn to parse physical space and visual inputs, hardware becomes just another peripheral to control. As Coogan joked: “At this point, horses remain one of the only major transportation platforms with no API, no OTAA updates, and no realistic path to MCP support.” Everything else with a digital connection is turning into an agent endpoint.

What to Do With This

Audit the three most annoying multi-step administrative workflows your team handles every week, such as vendor invoice approvals or local developer environment setups. Test an agent like Astra or Codex on one of those workflows this Thursday by giving it natural language screen instructions rather than building a custom script.