Agent Environments: A Survey of Synthesis and Verification

An agent environment provides a world in which an agent acts and a criterion for success. Earlier procedural generation and simulation automated parts of this object; recent digital-agent systems automate tasks, dynamics, and rewards together. This survey covers that literature across tool-use, web, computer-use, GUI/mobile, terminal, software, and embodied domains, with generative editing as an adjacent direction. Agent-environment synthesis is factorized into task generator, transition dynamics, and verifier, separating environment and dynamics validity, verifier validity, and trajectory success. Systems with synthesized dynamics use four substrates, while validation of task-success criteria is reported through heterogeneous protocols. Among 27 core systems, 23 expose a task-success mechanism; 18 of those 23 include an executable component, and 7 of the 23 report direct paired or human/reference evidence about evaluator outputs. A minimal reporting standard separates environment and verifier checks and, for generated rewards, tests known-positive and known-negative cases; three coded systems report this paired gate. Compositional environments for generative editing provide a further research direction.