AgentIn

AgentIn is the agent environment that [[KimiK3|Kimi K3]] opened according to 177: 详解Kimi K3：强到冲击Anthropic估值的模型什么样？. The source treats it as part of K3’s open infrastructure bundle alongside weights and low-level kernels, while still emphasizing that the full IO environment, task system, and raw expert checkpoints were not fully released.

The source-specific lesson is that agent capability and isolation are coupled. Zhao Chenyang / 赵晨阳 argues that Kimi’s approach is not simply to lock the model down, but to improve Agent Environment Isolation so the model can run with more useful permissions while failures remain contained.

AgentIn also connects training and deployment. The episode says K3 uses partial rollout so completed trajectories can be used for training before every long-running sample finishes, making Agent RL infrastructure more consistent with real agent execution.

Connections

• Kimi K3, Moonshot AI / 月之暗面, and Open Weight Release Boundary — model and release context.
• Agent Environment Isolation, Agent RL, Agent Post-Training, and Model Harness Co-Evolution — agent-training and execution context.
• AI Model Sandbox Escape, Frontier Model Cyber Misuse, and AI Cyber-Defense Utility — adjacent sandbox and safety context.