Environments determine what LLM-based agents can perceive, act on, and learn from — yet research on environment engineering has been fragmented. This survey argues that treating environments as first-class engineering artifacts (with modeling, synthesis, evaluation, and evolution stages) clarifies how agents acquire capabilities and how environments drive continual agent improvement.
Key Findings
- Systematic attribute-and-domain taxonomy: The paper organizes representative environments along eight attributes (e.g., observability, fidelity) and eight application domains, which helps compare environments beyond ad-hoc labels — so what? it makes trade-offs (complexity vs. controllability) explicit when selecting or designing an environment.
- Two synthesis paradigms: symbolic synthesis (rule- and template-driven) and neural synthesis (model-generated environments) are contrasted, along with evaluation methods for each. So what? practitioners can pick synthesis approaches aligned with repeatability (symbolic) or diversity/realism (neural).
- Environment evaluation and agent co-evolution: The survey describes evaluation metrics and frames agent evolution in four pathways (memory-, orchestration-, trajectory-, and exploration-centric). So what? this links environment design to measurable agent improvements and suggests targeted evaluation protocols.
- Three evolution paradigms: neural-driven, difficulty-driven, and scaling-driven environment evolution are identified, guiding how environments can be used to continually push agent capabilities. So what? it offers concrete experimental strategies for curriculum, synthetic data generation, and scaling studies.
Who it's for and trade-offs
Great fit if you design benchmarks, build training or evaluation platforms for LLM agents, or research agent–environment interactions and need a structured map of methods and metrics. Look elsewhere if you need low-level implementation recipes or reproduction-focused labs — the paper is a high-level, conceptual survey rather than a hands-on toolkit. It emphasizes taxonomy and research directions over turnkey code or datasets.
Where it fits
This work sits between benchmark papers that release concrete environments and methodological papers that focus on agent algorithms: it abstracts common design choices and evaluation needs across many existing environment efforts, helping unify evaluation and synthesis choices for future empirical work.
Methodological highlights
The survey’s practical takeaways include (1) matching synthesis paradigm to evaluation goals (repeatable vs. diverse scenarios), (2) treating environment attributes as explicit knobs when measuring agent generalization, and (3) planning environment evolution strategies (neural/difficulty/scaling) as part of long-term agent development rather than ad-hoc testbeds.