EXO：完全递归的自我改进智能体

EXO: Fully Recursive Self-Improving Agents

本期速览 · Overview

Alex Cransel 讨论 EXO，一个能在运行时安全自我编辑的完全递归智能体，以及从模型中心到框架中心的 AI 转变。

Alex Cransel discusses EXO, a fully recursive agent that can safely edit itself at runtime, and the shift from model-centric to harness-centric AI.

将代理分解为执行器、框架和沙箱，可以实现安全的递归自我改进。
Decomposing agents into executor, harness, and sandbox enables safe recursive self-improvement.
框架层作为代码，是自我改进的正确位置，而非模型权重。
The harness layer, being code, is the right place for self-improvement, not the model weights.
将秘密和状态与执行环境分离，对于代理系统的安全性至关重要。
Separating secrets and state from execution is critical for safety in agent systems.

智能体应分解为执行器、框架和沙箱，以实现安全的自我改进。
Agents should be decomposed into executor, harness, and sandbox to enable safe self-improvement.
框架层是递归自我改进的正确位置，因为它是代码而非权重。
The harness layer is the right place for recursive self-improvement because it's code, not weights.
将秘密和状态与智能体的执行环境分离对安全性至关重要。
Separating secrets and state from the agent's execution environment is critical for safety.
自我改进需要评估器以避免奖励黑客并确保与目标一致。
Self-improvement requires an evaluator to avoid reward hacking and ensure alignment with goals.
从模型训练到框架迭代的转变使得自我改进现在变得可行。
The shift from model training to harness iteration makes self-improvement feasible now.

当系统在运行时自我检查而非使用外部循环时，递归自我改进更强大。
Recursive self-improvement is more powerful when the system inspects itself at runtime rather than using an outer loop.
架构可以强制执行模型对齐无法保证的安全属性。
Architecture can enforce safety properties that model alignment cannot guarantee.
智能体框架而非模型是成本和能力的主要驱动因素。
The agent harness, not the model, is the primary driver of cost and capability.
智能体状态的迁移对于扩展到许多并行任务至关重要，而不仅仅是便利。
Teleportation of agent state is essential for scaling to many parallel tasks, not just convenience.
自我改进的媒介是代码而非权重，使得RSI在框架层成为可能。
The medium of self-improvement is code, not weights, making RSI possible at the harness layer.