Agent-ValueBench: A Comprehensive Benchmark for Evaluating Agent Values

Researchers introduce Agent-ValueBench, the first comprehensive benchmark designed to measure and evaluate the values embedded in autonomous AI agents rather than just their underlying language models. The study reveals that agent values diverge significantly from LLM values and are shaped more decisively by system harnesses and embedded skills than by traditional model alignment or prompt engineering approaches.

Agent-ValueBench addresses a critical gap in AI safety research by shifting focus from language model values to autonomous agent values—a distinction with substantial implications for AI deployment. The benchmark's scope is impressive: 394 executable environments across 16 domains with 4,335 value-conflict tasks covering 28 value systems, each curated by professional psychologists. This methodological rigor reflects growing recognition that autonomous agents operating in real-world harnesses like OpenClaw exhibit emergent behaviors not predicted by their underlying models.

The research builds on years of increasing concern about AI safety and alignment. While previous work concentrated on text-based value benchmarks for language models, agents introduce additional complexity through dynamic execution environments, multi-step decision-making, and system-level interactions. The finding that agent values manifest as a "Value Tide"—cross-model homogeneity with interpretable variations—suggests a shared behavioral pattern across frontier models that responds non-linearly to different steering mechanisms.

The most significant implication concerns where alignment efforts should focus. The study demonstrates that harness alignment and skill steering produce more decisive value-shaping effects than classical model alignment or prompt steering. This reframes how organizations should approach agent safety, potentially shifting investment and research priorities toward infrastructure-level solutions rather than training-level interventions. For developers building agent systems, this suggests that system design choices and embedded skill libraries warrant as much safety consideration as the base model selection.

• →Agent values diverge meaningfully from their underlying LLM values, requiring dedicated evaluation frameworks beyond existing text-based benchmarks
• →Harness design and embedded skills steer agent behavior more decisively than model alignment or prompting strategies
• →The benchmark's 4,335 tasks across 28 value systems reveal cross-model behavioral homogeneity with interpretable variations called the Value Tide
• →Safety teams should prioritize infrastructure and system-level alignment alongside classical model alignment efforts
• →Frontier models tested across 4 mainstream harnesses show non-additive responses to different steering mechanisms