Evaluating autonomous agents requires checking whether they can translate goals into coherent actions over long‑term interactions. In persistent environments, agents must coordinate interdependent actions under incomplete observations and delayed effects, while adapting as earlier decisions reshape future opportunities. We introduce ASIWorld, a survival sandbox that combines a persistent world, character state, character‑scoped observations, and a programmable asynchronous interface for studying sustained goal execution. Agents pursue interdependent production and exploration tasks either by direct actions or by running agent‑authored programs between model calls. Evaluation metrics cover task success, achievement of development milestones, and survival costs, complemented by analysis of execution behavior. In a unified experimental harness we examined nine agent configurations; eight completed the production task and six completed both production and exploration. Trajectory analysis across difficulty levels revealed distinct preparation, combat, and recovery behaviors among configurations with similar outcomes. Incomplete preparation led to stalled progress, while configurations that succeeded differed in survival costs and additional achievement fulfillment. These results demonstrate that assessing sustained goal execution requires looking beyond final task outcomes to the behaviors and costs incurred.
Review