Content

Problem

With over 26,000 resources available on the One Degree platform, keeping these resources up-to-date is a challenge. Volunteers and Resource Specialists at One Degree manually visit each resource’s website and compare details—such as description, website link, and steps to enroll—with information already stored on the One Degree platform. This process happens several times during the year, as timely, up-to-date information is crucial for the over 600 thousand users who rely on it to quickly find nearby social services.

Solution

We worked with One Degree to automate this update process through the use of an AI agent. Instead of volunteers manually visiting each website and comparing every data field, a One Degree Resource Specialist can run the AI agent with the desired resource IDs, which will initiate the AI agent to visit each resource website and generate suggestions of changes that need to be made to each resource. Resource Specialists can then review, edit, and approve each suggestion, decreasing the amount of time spent updating each resource.

Design

Users primarily consist of One Degree Resource Specialists, so a streamlined end-to-end user experience was central to our design process. Through early calls with One Degree, we learned that ideally, each run of the AI agent would process several hundred resources, which focused our efforts on presenting large volumes of suggestion data in a clear and manageable way. Much of our iteration centered on organizing suggestion data fields to support quick scanning and decision-making, as well as simple publishing of suggestions to ensure One Degree platform users have access to timely, up-to-date resource information. 

Tech Stack

We built the agent’s frontend using React and TailwindCSS for a clean, responsive interface. The backend, developed in Python with FastAPI, handled all routes and server logic, while LangChain powered the agent’s framework using ChatGPT’s LLMs. To enable web browsing capabilities, we integrated Playwright, allowing the agent to navigate and extract data from webpages. We used MySQL to store edits made by the agent before publication and implemented Google OAuth 2.0 for secure authentication. Finally, we connected to One Degree’s APIs to fetch resource data by ID and publish updates directly to their database.

Features

Agent Logs

To support transparency and oversight of the automated update process, the team implemented an agent logging system that provides real-time visibility into the AI agent’s activity. As the agent runs, logs are streamed to display its current state and actions, including the web pages it navigates, any errors encountered, and other operational details throughout the process. These logs allow Resource Specialists to monitor the agent’s progress and better understand how update suggestions are generated, helping ensure the system remains reliable while aligning with the broader goal of efficiently maintaining accurate resource information on the One Degree platform.

Custom Runs

To give Resource Specialists greater control over how updates are generated, the team developed a system for initiating customizable agent runs. One Degree administrators can start and rename individual runs of the AI agent, allowing them to organize and track different update sessions. Each run can also be configured to target specific resource IDs and selected data fields, enabling specialists to focus the agent’s work on particular resources or types of information. This flexibility helps streamline the review process and ensures the automated workflow fits naturally into One Degree’s existing resource maintenance practices.

Agent Suggestions

For each resource included in a run, the AI agent generates either a suggested update or an error message based on the information it is able to retrieve. Successful suggestions include a confidence score that indicates how certain the agent is that it has identified the correct information from the resource’s website. Alongside this score, the agent outputs a structured JSON object containing the proposed updates, formatted to mirror One Degree’s existing database schema. If the agent encounters issues—such as invalid resource IDs, closed resources, or other errors during navigation—it generates an error entry instead. This structured output allows Resource Specialists to efficiently review the agent’s findings while maintaining consistency with One Degree’s data model.

Diff Viewer

To support careful review before updates are applied, the system includes a JSON diff viewer for each suggestion generated by the AI agent. The viewer highlights differences between the original resource data stored in One Degree’s database and the modified JSON proposed by the agent, allowing Resource Specialists to quickly identify and evaluate changes. Within this interface, users can edit the suggested values, revert individual changes, or approve the update. Once approved, the revised information is written to One Degree’s database and reflected on the platform. Each suggestion also includes source links to the webpages the agent referenced while gathering information, providing additional context and traceability during the review process.