这篇论文提出了一种让智能体跨平台迁移的架构，解决了智能体长期运行时的身份连续性问题。
该论文提出了一种智能体架构，使智能体能够超越其模型、工具和宿主的生命周期。研究将智能体分为两部分：持久存在的核心部分（包含身份、私有内存和代码版本历史）和可替换的基础设施部分（推理模型、运行工具、服务器和用户接口）。作者通过833个核心测试及92个提供商和库测试验证了架构可行性，并在实际部署中成功实现了模型版本、接口和物理主机的迁移。
What a super interesting paper this one is. They propose an architecture for agents that outlive th...
What a super interesting paper this one is. They propose an architecture for agents that outlive their model, harness and host. Today we describe an agent by whatever model and harness it happens to run on. That works for a single session. It says very little about an agent that runs for months and gets moved to a new model, a new harness, or a new machine along the way. The paper splits an agent in two. One half is the agent itself, and it persists. Its identity, its private memory, and its own code with version history. The other half is plumbing you can replace. The model doing the reasoning, the harness running it, the server hosting it, and the ways people reach it such as chat, an API, or a UI. Swap the plumbing and you have moved the agent rather than built a new one, as long as the handoff is authorized and keeps the record of where it came from. The handoff is six steps. Pause the agent, save its state, check the save is valid, attach it to the new setup, load the state back, then let it run again. They ran the frozen public release on a clean machine and it passed 833 core tests plus 92 more for providers and libraries. They also swapped model versions, interfaces and physical hosts on live deployments. The authors are careful about what this proves. It shows you can move an agent without breaking it mechanically. Whether the agent still behaves like itself afterwards is a separate question. Paper: arxiv.org/abs/2609.00546 Chat with Paper: academy.dair.ai/papers/runtime… 💬 6 🔄 5 ❤️ 21 👀 1749 📊 11 ⚡