向人工智能代理委派工作时，应当根据两个关键因素来决定自主权的程度[1]。这两个因素分别是：代理工作是否易于验证，以及撤销其错误的成本是否低廉[1]。
基于这两个维度，可以构建一套四级委派框架[1]。最低级别（Level 0）适用于既难以验证又难以纠正的任务，此时代理仅充当辅助工具[1]。当错误容易纠正但难以验证时为Level 1，验证容易但纠正困难时为Level 2，而最高级别（Level 3）则是验证和纠正都相对便利的场景[1]。在PostHog重写SQL解析器的实践中，工程师Robbie采用了Level 2的策略——利用自动化机器进行代码验证，同时在生产环境中通过影子模式和分阶段切换作为最终的安全保障[1]。相比之下，PostHog的Scouts代理已达到Level 3水平，可以按计划自动运行、调查产品数据信号并起草拉取请求[1]。
为了在各个委派级别上提升代理的能力，应当采取多项措施[1]。这些措施包括明确定义可量化的目标、利用大语言模型进行评判、编写定制化技能模块、训练特定领域的专用模型，以及建立覆盖领域专家知识的上下文库[1]。
A framework for determining appropriate autonomy levels when assigning tasks to AI agents has been outlined, based on two critical factors: whether an agent's work can be easily verified, and whether mistakes can be cheaply reversed [1]. This decision-making structure establishes four distinct delegation levels, ranging from agents functioning as assistants with minimal autonomy to fully autonomous operation [1].
The framework categorizes delegation into four tiers according to verification difficulty and error reversal costs [1]. Level 0 applies when verification is difficult and reversal is expensive, limiting agent autonomy significantly. Level 1 operates in scenarios where verification remains challenging but errors are inexpensive to undo. Level 2 represents situations with straightforward verification but costly reversals, while Level 3 encompasses cases where both verification and reversal are simple, permitting maximum autonomy [1].
PostHog's implementation of this framework demonstrates practical application across delegation levels. At Level 2, engineer Robbie employed automated machine verification when rewriting the SQL parser, but maintained additional safeguards by deploying the solution in shadow mode with staged production rollout to serve as final safety checks [1]. At the highest autonomy level, PostHog developed Scouts agents that operate on a schedule, analyze product data signals, and automatically draft pull requests [1]. To optimize agent effectiveness across delegation levels, organizations are advised to establish clearly defined measurable objectives, leverage language models as evaluators, develop custom skills, train domain-specific models, and build repositories of expert-level context [1].