Workspace-Bench evaluates how well AI agents operate in realistic workspaces with large-scale file dependencies. It measures whether agents can discover relevant files, reason across implicit context, modify artifacts safely, and deliver the right output under rubric-based evaluation.
Independent benchmark summaries for workspace agents, model backbones, and benchmark composition.
Workspace-Bench reports a clear gap between human performance and current evaluated agents.
Major milestones and feature additions since the benchmark's public launch.
Workspace-Bench and Workspace-Bench-Lite were released together with the public repository, benchmark structure, task metadata, rubrics, and file-dependency information.
The official paper introduced the benchmark design, workspace-learning evaluation dimensions, scoring protocol, Lite split, and public leaderboard framing.
The public Lite leaderboard was released with agent/model rankings, rubric pass rates, threshold markers, and distribution figures aligned with the released benchmark artifacts.
Jump into the official leaderboard, dataset composition, and scoring methodology from a single entry section.
Browse public Lite systems, full-summary rows, threshold views, and capability-oriented leaderboard analysis.
Inspect worker profiles, collaboration types, rubric distributions, file dependency coverage, and task metadata structure.
Review the rubric-based scoring process, submission requirements, and how benchmark evidence is structured.
The official Workspace-Bench framework figure summarizes how role-specific workspaces, benchmark tasks, agent execution, and capability-oriented evaluation fit together.
The benchmark focuses on workspace learning: the ability to identify explicit and implicit file dependencies, integrate context across artifacts, choose the right actions, and produce dependable deliverables in large workspaces.
Navigating deeply nested directory structures and identifying relevant files from noisy candidates.
Finding files that provide essential context, references, and domain knowledge needed to complete a task.
Aggregating result files that contain required outputs, formats, and baseline information.
Tracing explicit references, semantic connections, and contextual links between related documents.
Connecting information across diverse modalities including documents, spreadsheets, presentations, and code.
Understanding file versions, revisions, and derivation relationships (e.g., report_v1, report_final).
Workspace-Bench exposes a large gap between human performance and the best AI agents. The gap grows in profiles with more hidden dependencies, denser file graphs, and stricter rubric requirements.
Use the benchmark website as a compact index to the paper, repository, dataset composition, and submission process.
Browse official scores, cost comparisons, and capability views.
Inspect task composition, file types, workspace sizes, and profile distributions.
Read the workspace learning definition, scoring rules, and reproducibility protocol.
See concrete examples of hidden dependencies and rubric-based grading.