Researchers presented the EconSkills library for training AI agents

Researchers introduced EconSkills — a specialized library and framework for evaluating AI agents' skills when working with dynamic economic data. The system converts successful interaction trajectories with web resources from the EconWebArena dataset into parameterized standard operating procedures.

Each skill in the library includes an area of application, a step-by-step navigation algorithm, specific instructions for specific sites, verification mechanisms, and error recovery protocols. In the process of abstraction, the original values from real requests are replaced with universal placeholders, allowing agents to transfer the experience gained to new, previously unseen tasks.

Testing showed that using EconSkills significantly increases the likelihood of success compared to methods without pre-training. The application of abstract procedures proved to be more effective than direct reproduction of "raw" trajectories, requiring the model to take fewer steps to achieve the goal. When scaling the library, the skill search mechanism demonstrates competitive results, especially on tasks directly covered by the database.

The development is aimed at solving the problem of agents "forgetting" accumulated experience when switching between different web resources. The authors emphasize that the proposed approach allows for accurately determining when an agent needs a procedural hint for navigation, and when semantic data verification is required. The project is distributed under the arXiv license for non-exclusive distribution.