Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments
Terminal-Universe is a framework that reconstructs reusable terminal environments from agent trajectories by replaying file operations and using a completion agent to supply missing files and dependencies. It then reconstructs the original intent task and synthesizes new tasks, scaling them along two dimensions.
As terminal-based code agents become prevalent, agent trajectories have accumulated at scale, while realistic, executable environments remain scarce. Terminal-Universe turns each trajectory into a reusable environment by replaying recorded file operations to restore files before modification, then using a completion agent to fill in missing files and dependencies. On the recovered workspace, it reconstructs the original intent task and synthesizes entirely new ones, scaling tasks along two dimensions.
The approach leverages tool-execution history in trajectories to expose environment structure and contents, enabling reconstruction without generating environments from scratch. Replaying file operations yields a partial workspace, and a completion agent supplies missing components. This allows each trajectory to be converted into a queryable environment that provides execution feedback, supporting verifiable task synthesis and continued interactions.
This framework addresses the scarcity of realistic, executable environments for agent post-training by repurposing existing trajectory data. It could reduce the cost and effort of environment creation, potentially accelerating development of terminal-based code agents and enabling more scalable training and evaluation pipelines.
Terminal-Universe could lower barriers to creating training and evaluation environments for code agents, potentially improving agent performance and reducing development costs for companies building terminal-based AI systems.
Observable next signals include publication of implementation details, release of code or datasets, and adoption by agent training frameworks. Further research may explore scaling task synthesis along additional dimensions and applying the method to other tool-execution domains beyond terminals.