Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 18·5 min read
StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling
TL;DR
StateM is a novel runtime that enhances long-horizon agents' performance by organizing execution around durable states and recoverable runbooks, achieving a raw accuracy of 95.3% on Terminal-Bench 2.1.
✦ Why It Matters
Implement StateM to enhance your AI agents' performance on complex tasks without retraining models.
Key Takeaways
- Implement StateM to improve long-horizon agent performance.
- Utilize recoverable runbooks for better task execution.
- Achieve high accuracy with minimal adaptation costs.
- Leverage family-specific runbooks for targeted performance gains.
How It Works
StateM organizes agent execution around durable states and phase-local context, allowing agents to maintain relevant information throughout long tasks. It employs checked transitions and recoverable runbooks, which are structured procedures that agents can refer back to, ensuring they do not skip important steps or lose track of learned lessons.