VibeLifeBench: Can Your Life Agent Be Proactive and Persistent in a Living World?
Xiaohongshu Inc
cs.CL, cs.AI
2026-08-11
VibeLifeBench is a long-horizon life-domain agent benchmark: 200 tasks across ten domains, multi-week simulated horizons, 22 mock services, and a world that advances on its own clock with about 70 percent of events not user-triggered. The strongest of seven frontier models, Claude Opus 5, averages only 32.5 of 100, and none is competent across all domains.
Most agent benchmarks use short, self-contained requests in static environments: a clear instruction, a passive one-shot execution, a world that changes only when the agent acts. Everyday life assistance is nothing like that. A task runs for weeks rather than minutes; the world keeps changing while the agent is not being prompted; many constraints (passport validity, a budget cap, a phishing email) are never stated outright. An agent that only answers the request in front of it will fail. What is needed is an agent that stays proactive and persistent: deciding on its own when to act, ask, or stay silent, noticing changes nobody announced, and keeping one plan coherent from the first day to the last. VibeLifeBench measures exactly this.
The benchmark has 200 tasks spread evenly over ten everyday-life domains (travel, finance, litigation, renovation, career, fitness, exam preparation, rental, shopping, team building), 20 each. Each task is a scripted multi-week timeline running on 22 mock services (banking, credit card, brokerage, flight, hotel, rail, and car booking, maps, email, calendar, notes, a notification hub, and more).
Several design choices matter. The world advances on its own clock: about 69.9 percent of 7,453 events are environment-driven and only 30.1 percent are user messages; 1,483 mutations trigger no agent turn and alter state silently, so only an agent that re-inspects the world catches them. The median simulated horizon is 29 days (longest about 111), with a median of 24 stages and 7 services per task; all 22 services are exercised. Scoring reads only observable artifacts and never the agent's internal reasoning, across 12,261 weighted checks (median 58 per task) covering end state, timeliness, and whether implicit constraints held. Cross-stage and final checks are 19.1 percent of items but 26.8 percent of weight, so a single safety or hardening failure costs more than many routine subtasks. Each task runs three times, reported as avg@3, max@3, min@3, and within-task standard deviation.
All seven frontier models (Claude Opus 5, GPT-5.5, Gemini 3.5 Flash, Claude Opus 4.8, GLM-5.2, Kimi-K2.6, DeepSeek-V4-Pro) score low. Claude Opus 5 leads at avg@3 32.5 (max@3 41.2, min@3 23.8), GPT-5.5 is 30.1, Gemini 3.5 Flash 27.5, and DeepSeek-V4-Pro 21.1. The seven sit in a narrow 21-to-33 band, far closer to each other than to competence, and larger or newer models do not clear the bar. Every model's min@3 is at most 23.8 and within-task standard deviation reaches 10.0, so even a correct run is hard to reproduce.
By capability axis, proactivity and persistence are the weakest, no model exceeding 33.6. Persistence and bookkeeping accounts for 22.2 percent of all failed checks, remarkably stable at 22.0 to 23.2 percent for every model; models do write, but rarely the specific cross-stage, auditable, well-formed artifacts the criteria require. Per-stage pass rate in the last third of the timeline is 10 to 15 points below the first third for every model, so long-horizon coherence decays over time.
This puts three properties existing benchmarks overlook (proactivity, living-world adaptation, long-horizon coherence) into one measurement, and shows that strong tool-calling in professional settings does not automatically transfer to managing weeks of life errands. Spending more tokens does not buy a higher score (Gemini 3.5 Flash reads the most context and takes the most turns yet lands mid-pack); what matters is whether state is persisted durably and structurally. For teams building personal assistants or long-horizon agents, the benchmark and its four named directions (durable bookkeeping, proactive re-perception, safety hardening, suppressing time decay) are a direct roadmap.
Scoring is deterministic and reads only observable artifacts, more reliable than LLM-as-a-judge, but the capability axes are assigned by keyword matching over check names, so they are indicative rather than exact, and 41.8 percent of failures fall outside the named categories. The 22 services are mocks, not real integrations, so the benchmark covers mechanisms rather than the long tail of real API trouble. Tasks and implicit constraints are hand-authored with cross-validation, but 200 tasks is a limited scale, and generalization to real user life streams is not validated. All seven models use the same native tool-calling scaffold at their strongest reasoning setting, deliberately excluding scaffold tuning, so scores reflect the model rather than best engineering.