ASTRA-bench: State-of-the-Art Agents Degrade Significantly Under High-Complexity Personal Context — Argument Generation Is the Primary Bottleneck

arXiv·medium signal

ASTRA-bench (arXiv:2603.01357) is a new evaluation that tests tool-use agents on 2,413 scenarios grounded in time-evolving personal user context — the kind of longitudinal, multi-party life data real assistants must handle — with an interactive toolbox and complex multi-step intents. Evaluation of Claude Opus 4.5 and DeepSeek-V3.2 reveals significant performance degradation under high-complexity conditions, with argument generation (deciding which arguments to pass to which tools in what order) identified as the primary failure mode. Most existing benchmarks are context-free and single-turn; ASTRA-bench is the first to combine dynamic personal context with tool-chaining and multi-step reasoning at this scale.