Thin Harness, Strong Contracts: Production-Oriented Agent Harnesses for Stateful AI Agents

Key Points

• The paper aims to promote production-oriented agent harnesses that optimize agent cognition while maintaining robust boundaries.
• Introduces a comprehensive artifact package including manuscript, scripts, and benchmark cases.
• Defines system boundaries for stateful AI through tools like permissions and audit systems.
• Utilizes mechanisms for replay, memory scoping, and human review in agent interactions.
• Demonstrates improved control over tool calls and permissions in the AI system.
• Establishes reproducibility through extensive logs and documented benchmarks.
• Validates the framework with evidence snapshots and regenerated result tables.