Wes Roth tests OpenAI Astra across long-running computer-use tasksA long-running AI agent maintains task state and continues observing, deciding, and acting across an extended period. after receiving access to the model. He has it create game assetsGame generation uses AI to create or assemble playable assets, scenes, rules, audio and interactions., build and animate objects in Blender, assemble projects in Unreal Engine and Unity, add generated audio and play through the result to check whether the experience worksAgent evaluation tests whether an AI agent completes tasks correctly, consistently, and within its required boundaries..
Roth says the same system created a controller for RimWorld, ordered groceries through Instacart and edited a video by removing pauses. He also demonstrates voice interaction inside an active project and argues that these workflows go beyond ordinary code generation because the model can move between applications and evaluate its own output.
The strongest game was still a rough first pass with basic level design, but Roth considers the result unusually complete for roughly two hours of autonomous work. He also warns that agents operating overnight can reach browsers and sensitive systems, so he removed stored profiles and restricted accessAn agent permission boundary limits the information, tools and actions an AI agent can use during a task. before leaving them running.
Watch on YouTube
