Web Agent

green agent evaluates WebShop shopping tasks in a text‑only Gym environment. It orchestrates episodes by resetting the environment, sending observations to the purple agent, executing returned actions (search/click/buy), and collecting programmatic rewards. It reports structured JSON artifacts containing total reward, success, and per‑step traces. This provides a reproducible benchmark for instruction following in e‑commerce search and product selection without an LLM judge.

Our green-agent evaluates how well a white agent can understand and predict user shopping behavior in the context of online grocery shopping. The green-agent sets up a test in which white agents will be given both a user’s past purchases and the documentation for a shopping API, and white agents will have to use said shopping API to build the best basket for the shopper given the context. Ground truth will be measured against what the users ultimately purchased (as derived from the real transaction dataset). We built a green agent to test how well white agents can auto-shop for your groceries given previous purchases. We will provide an agent with a partial transaction history for a given user which contains their last n shopping trips and provide said agent with an e-commerce API (built in house on training data) so they can make searches, view results, and build a basket. When the agent is done building said users' n+1 basket, we check and see what % of items they predicted which the user actually checked out (since we have the users’ complete transaction history).

This Green Agent defines a deterministic and fully offline benchmark for evaluating agentic systems that retrieve paginate deduplicate and normalize Comtrade style international trade data. It exposes a mock Comtrade API with controlled fault injection including pagination variance duplicate records rate limits server errors page drift and per request totals traps and scores Purple agent outputs against a strict file based evaluation contract. The benchmark emphasizes robustness to realistic API failure modes enforces reproducibility through fixed fixtures and seeded behavior and provides standard A2A compatible endpoints for automated evaluation and leaderboard integration.