AI agents are increasingly acting as digital coworkers that can browse the web, write code, run experiments, and assist humans over long horizons. Existing benchmarks, however, largely rely on synthetic data or isolated tasks and miss the complexity of real knowledge work, especially the multi‑party dynamics where critical information is scattered across conversations, code, and evolving decisions. Even a simple workplace query may require locating the right channel or direct message, recovering context from lengthy discussions, inspecting repository state, reconciling stale or conflicting evidence, and synthesizing across sources.\n\nTo address this gap we introduce SlackBench, a benchmark built on privacy‑preserving reenactments of real academic research projects. It provides reproducible environments with multi‑party dialogue, evolving decisions, and corresponding GitHub code and Google Docs artifacts. The dataset contains 140 queries covering heterogeneous‑source search, abstention on unanswerable or false‑premise questions, goal‑directed summarization, and circumstance inference over long conversations. These tasks cannot be reduced to static corpus QA; agents must decide which conversations matter, traverse threads, inspect repository state, and synthesize evidence across communication and code.\n\nAcross evaluations of 33 frontier models and agent harnesses, the best system reaches about 80% while most score below 66%, highlighting substantial room for improvement. SlackBench thus offers an extensible framework for assessing agents in realistic collaborative work environments. The dataset is available at https://huggingface.co/datasets/HuanzhiMao/SlackBench.\n\nReview: SlackBench fills a critical gap in benchmarking real‑world collaborative scenarios, giving researchers a systematic and reproducible platform. Future extensions could broaden task diversity and cover additional industry domains.