Why do AI agents fail at workplace social interaction?

TheAgentCompany creates a self-contained environment simulating a small software company — web interfaces, code repositories, communication platforms, and simulated colleagues. Tasks span multiple job categories: browsing the web, writing code, running programs, and communicating with coworkers. The most competitive agent completes 30% of tasks autonomously.

The failure pattern is revealing. Three categories are specifically hardest:

1. Social interaction — tasks requiring communication with simulated colleagues, asking for information, and coordinating outputs. This is consistent with Why do reasoning models fail at theory of mind tasks? and Why do reasoning models struggle with theory of mind tasks? — formal AI reasoning capability does not transfer to social contexts.

2. Complex professional UI navigation — professional tools designed for human workflows (not API access) require sequential multi-step interactions where each step builds context. This connects to Are reasoning model collapses really failures of reasoning? — the execution layer, not the reasoning layer, is the bottleneck.

3. Private knowledge domains — tasks where publicly available resources don't exist, requiring domain-specific understanding of internal processes and conventions.

The benchmark design captures something most agent benchmarks miss: real workplace tasks require interaction — asking colleagues for information, sharing partial results, negotiating task requirements. Since Why can't advanced AI models take initiative in conversation? documents that current agents can't lead conversations, and since When should AI agents ask users instead of just searching?, the social interaction gap is both the largest and the least addressed.

The 30% figure provides a calibration anchor: simpler tasks are automatable, but the remaining 70% requires capabilities that scale differently from raw reasoning performance.

Enterprise benchmark convergence: CRMArena-Pro (CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions) extends this to enterprise CRM settings with 19 expert-validated tasks across customer sales, service, and configure-price-quote scenarios. Leading agents achieve approximately 58% single-turn success rate — but drop to 35% in multi-turn settings. Workflow Execution is the tractable outlier (83%+), while other business skills present greater challenges. Most critically, agents exhibit near-zero inherent confidentiality awareness — improvable with prompting but at a cost to task performance. The single-turn → multi-turn drop (58% → 35%) is consistent with Why do language models lose performance in longer conversations?, and the 35% multi-turn figure converges with TheAgentCompany's 30%, suggesting a stable performance ceiling for current agents in realistic workplace settings.