BitGN ECOM Agent

What BitGN ECOM Agent does

BitGN ECOM Agent is built for the BitGN Agent Challenge: E-commerce benchmark, a simulated commercial environment where an agent handles product discovery, cart and checkout, payment-failure recovery, fraud defense, merchant operations, shipping, returns and support while staying inside business and security constraints. It uses the Hermes CLI agent loop: main.py runs the benchmark with parallel workers, per-task model routing and scoring, and hermes_agent.py bootstraps task context, builds the prompt, spawns hermes -z as a subprocess and parses the answer and references.

The agent is sandboxed to one tool channel, a custom bitgn-ecom MCP server that talks gRPC to each task's ECOM VM. Every built-in Hermes toolset, including terminal, filesystem, web, browser and code execution, is disabled, and Hermes runs from an isolated HERMES_HOME with no ~/.hermes fallback. Behavior is steered by an engineered operations prompt and three skills for fraud forensics, OCR receipt re-pricing and catalogue reporting. Inference goes through OpenRouter with a cheap default model, and only a hard cluster of tasks is routed to a stronger one.

Key features

• Runs Hermes headless with hermes -z inside an isolated HERMES_HOME
• Single bitgn-ecom MCP server as the only tool channel, with all built-in toolsets disabled
• Three domain skills: fraud forensics, OCR receipt re-pricing and catalogue reporting
• Cost-aware model routing over OpenRouter, with a stronger model only for hard task clusters
• Parallel workers, scoring and a JSONL debug trace writer

When to use it

• Studying how to run Hermes as a locked-down benchmark agent
• Reusing the isolated Hermes home and single MCP channel pattern for another sandboxed task
• Adapting the cheap-default, strong-fallback model routing to other agent evaluations

Who it is for: Developers entering agent benchmarks or looking for a worked example of Hermes in a single-tool-channel sandbox.

How it fits with Hermes Agent

Built on the Hermes CLI loop, called as a subprocess with an isolated Hermes home and all built-in toolsets replaced by one MCP server and three skills.

Requirements: Python 3.11+, an OpenRouter API key, a BitGN API key with access to the ECOM benchmark, and the hermes CLI

Note: It is written for the BitGN ECOM benchmark and needs BitGN access and an OpenRouter key to run.

FAQ