Taking the PULSE of Enterprise Workflows: Evaluating Persistent Project-State Understanding

Enterprise agents must reason over work histories that are heterogeneous, evolving, and only partially visible to different collaborators. We formulate this challenge as persistent project-state understanding and introduce PULSE, a benchmark with questions spanning open loops and blockers, project evolution, and team-versus-individual perspectives. PULSE provides deterministic, time-indexed project-state ground truth, enabling us to separate failures in evidence retrieval from failures in reconstructing state from evidence. We find that PULSE is quite challenging even for frontier models and SOTA retrieval systems. Across representative retrieval, long-context, and oracle settings, we observe that retrieval explains much of the performance gap, but explicit project state still yields substantial gains; and perspective-conditioned questions remain most difficult even given full knowledge of project state.