ComPASS: Towards Personalized Agentic Social Support via Tool-Augmented Companionship

The meaningful distinction in ComPASS is not that it uses tools, but that it grounds tool selection in established social support theory, meaning the agent is supposed to know when to search for resources, when to reflect back emotion, and when to do nothing, rather than defaulting to a single response mode. ComPASS-Bench is the part worth scrutinizing: benchmarks for emotional support are notoriously hard to validate because ground truth is contested even among clinicians.

The agentic framing here connects directly to the broader agent benchmarking wave we covered around mid-April. CoopEval (covered April 16) raised a pointed question about whether LLM agents behave appropriately in socially consequential situations, finding that models defect rather than cooperate under pressure. ComPASS is asking a related but distinct question: can agents calibrate behavior to individual emotional needs over time? The social stakes in companionship are arguably higher than in game-theoretic dilemmas, yet the evaluation infrastructure is less mature. The Codex and MM-WebAgent coverage from the same period shows tool use becoming a baseline expectation for agents, which makes the psychological grounding in ComPASS the actual differentiator worth examining.

Watch whether ComPASS-Bench gets adopted by any third-party research group within six months. If it does, the benchmark has legs; if citations cluster only around the original authors, the evaluation criteria likely reflect design choices too specific to generalize.