As large language models (LLMs) increasingly operate as agents—working within environments that provide different tools, permissions, data schemas, and forms of interaction—it is important to better understand how the structure and design of a given workspace might shape an agent’s performance and behavior. While interest in this relationship is growing, relevant work remains scattered across communities, siloing insights and narrowing the scope of inquiry. This exploratory research agenda seeks to address that fragmentation by treating an agent’s working environment as a distinct object of study, drawing together perspectives from otherwise separate disciplines to better define what constitutes an agent’s workspace and how differences across workspaces constrain or enable the work an agent can perform.
This agentic shift has substantially expanded what a given language model can reliably accomplish, allowing them to better reason through a task, take actions, evaluate results, and modify their work as needed. Coding agents such as Claude Code and OpenAI’s Codex have been particularly successful examples of this development. Yet while all software is written in code—making it tempting to assume agents could be harnessed and transferred to to similarly succeed at other jobs that use computers—not all digital workspaces are built or bundled the same. What we can expect to find in a programmer's working environment (e.g. executable artifacts, version histories, dependency structures, error messages, tests, and reversible actions) may not exist in comparable forms across occupations. If the ability of coding agents, and LLM agents more generally, to successfully complete the work given to them depends not only on their underlying model's trained capabilities but also on how a given workspace allows those capabilities to be reliably enacted then a model-centric understanding of AI is necessarily incomplete. What an agent can actually accomplish must instead be understood relationally, as an outcome of the interaction among the model, the work it is asked to perform, and the workspace through which that work becomes perceptible, actionable, and verifiable.
To develop a more ecological account of agent capability, this research agenda aims to draw on technical fields such as reinforcement learning, control theory, and multi-agent systems; fields concerned with interaction and cognitive work, including human–computer interaction, human factors, cognitive systems engineering, and embodied and distributed cognition; fields concerned with work and institutions, including the sociology of work, organization studies, labor economics, and science and technology studies; and philosophical traditions such as phenomenology and the philosophy of technology. Together, these traditions offer complementary ways of understanding how an actor’s capacities are expressed, constrained, and transformed through its relations with tools, representations, feedback loops, work practices, other actors, and institutions. Applied to LLM agents, this approach could: help researchers design more representative benchmarks; help product, engineering, and user-experience teams identify when performance depends on workspace and systems design rather than further model development; and help economists and policymakers distinguish technical potential from practical deployability and realized impact.
