Brandon Thomas of ArtCraft describes agent-built creative apps and the operational bottlenecks behind them in this 20-minute AI Engineer talk.

Build familiar tools with programmable cores

  • Thomas says a large prompt produced a functional Photoshop approximation in roughly 24 hours.
  • The prompt separated interface work, tool behavior, data structures, file importers and automation interfaces.
  • Familiar layouts preserve artists’ muscle memory; command-line interfaces and MCP, the standard for connecting agents to tools, expose the same apps to automation.
  • Artists can adapt open-source tools to narrow studio workflows instead of waiting for a vendor’s roadmap.

Rust helps the apps; compilation slows the factory

  • Thomas credits native Rust with startup speed, cross-platform builds and browser deployment through WebAssembly.
  • The team is extracting reusable interface, 2D and 3D components for community projects.
  • He identifies compile times, permissions and overloaded continuous-integration infrastructure as practical constraints.
  • His recommendation keeps humans close to backend schemas and API design while allowing more autonomy on isolated desktop-app work.

Autocraft is a proposed next layer

  • Thomas describes a forthcoming open-source factory for issue intake, triage, coding, evaluations and post-training from recorded decisions.
  • He also reports large community and download figures; these are his account, without independent measurements in the talk.
  • Caveat: the 24-hour result was an approximation. Thomas explicitly disclaims full Photoshop parity, and ArtCraft’s current catalog labels these tools early alpha or in development.

“Let’s own our apps, our infra, our protocols, our models, all of it.”