How woooodyy/agentgym works
• Controller / Agent Manager: Represented by the “agentenv/agentenv/controller” package and core module (which includes agent, environment, task, and utility functions). This is the central hub that evaluates agents, connects to various environment servers, collects data, and performs training.
• Environment Services: Show the multiple environment implementations. Each “agentenv-...” folder (e.g., agentenv-webshop, agentenv-webarena, agentenv-alfworld, etc.) provides a standardized HTTP API with endpoints such as /createEnv, /observation, /available_actions, /step, and /reset. Each such environment server should be drawn as an independent microservice or container.
• EnvClients: These are the client-side facades that wrap the HTTP communication with environment servers so that the AgentController can call them in a uniform manner.
• Training and Evaluation Modules: Separate components for training (AgentEvol, behavioral cloning trainers, dpo trainers) and evaluation—this includes benchmark pipelines that use AgentEval and data collection from AgentTraj.
• Utility and Logging Services: Include the utility modules (dispersed under utils/ directories), logging, and distributed evaluation components.
Relationships and Interactions
• Data flow from the controller to “Training Modules” (trainer components in agentenv/agentenv/trainer) as well as the data flow toward “Benchmark Modules” that interact with external datasets (AgentEval, AgentTraj).
• That utilities, logging, and error handling modules “support” both the agent controller and environment handlers.
Architectural Patterns
• Microservice Architecture: The environments being separate services (each packaged as its own “agentenv-*” module) demonstrate a microservice or modular architecture.
• API-centric Communication: Emphasize the uniform interface (HTTP APIs such as /step, /reset, etc.) that governs interactions between agents (via EnvClients) and the different environments.
• Separation of Concerns: Clearly separate the domains—environment simulation, agent control, training, evaluation, data collection, and utility functions.