Researching how agents learn, act, collaborate, and improve.
RSIGym is an agent-native research environment built on Everything as a Service, exposing training, inference, rollout, evaluation, and sandbox execution through reusable services. It supports data, harness, and joint improvement under shared budget and permission controls, and introduces RSI-Index to measure how effectively frontier research agents improve models across five benchmarks.
Gödel Forest organizes data-centric recursive self-improvement as an ensemble of co-evolving search trees. Each agent grows a persistent tree of training-data strategies, while shared memory distills successes and failures into compact lessons that guide exploration across the forest. Evaluated on RSIBench-Data across six domains, it outperforms a single-agent baseline and reduces wall-clock time on five tasks, showing how collective learning can balance deep research with broad exploration.
Personal assistants must operate in a world that keeps changing, often for weeks and without continuous user prompts. We introduce VibeLifeBench, a benchmark of 200 long-horizon tasks across ten everyday-life domains and 22 simulated services. Seven frontier models all perform poorly, revealing a fundamental gap between agents that respond to requests and agents that can proactively monitor change, preserve intent, and act consistently over time.
Can agents turn evidence of model failures into better training data and improved models? We introduce RSIBench-Data, a controlled benchmark that isolates data-centric research decisions within a fixed post-training stack. Across six benchmarks, agents show that they can form useful hypotheses and improve initial attempts, but struggle to preserve strong results and convert feedback into consistent progress—highlighting a central bottleneck in recursive self-improvement.
What happens when an agent’s reusable skill library grows from dozens of skills to hundreds or thousands? Across 15 frontier models, 1,141 real-world skills, and more than three million routing and execution decisions, we identify coupled scaling laws governing skill selection and task execution. The results show that agent performance depends not only on model capability, but also on how skills are structured, differentiated, and exposed.
Coding agents increasingly operate on shells, repositories, and user files, making least-privilege authorization essential for safe deployment. We introduce AuthBench, a benchmark of 120 realistic terminal tasks with human-reviewed permission boundaries and executable security checks. Current models routinely grant unnecessary access while omitting permissions required for execution, showing that authorization cannot be solved by simply making agents more conservative or giving them more reasoning time.
Coworker agents must complete workflows that span multiple days while emails, calendars, documents, and multimodal evidence continue to change around them. We introduce ClawMark, a benchmark of 100 tasks across 13 professional scenarios, evaluated through stateful sandboxed services and deterministic rule-based checks. Frontier agents often make partial progress but rarely complete the full workflow, with performance dropping sharply after the environment changes.