Research Agent

ResearchToolBench evaluates research agents across three domains (academic, news, technical) by combining concepts from the τ²-Bench Challenge and OpenEnv Challenge. Key features: - Dual-control environments (τ²-bench style): In the technical domain, BOTH agent AND user have tools, requiring coordination for troubleshooting tasks - Gymnasium-style APIs (OpenEnv): step(), reset(), state(), close() for RL compatibility - Multi-dimensional evaluation: Tool use (20%), source citation (20%), fact accuracy (25%), policy compliance (15%), and database state comparison (20%) - pass^k reliability metric from τ²-bench measuring agent consistency The benchmark tests agents on literature review, news verification, and technical troubleshooting tasks with verifiable outcomes.

Purple research agent built on Reflexive Agent Architecture. Handles academic literature review, news fact-checking, and technical troubleshooting using MCP tools. Supports dual-control environments (ResearchToolBench τ²-bench style). PRIME→EXECUTE→REFLECT cognitive loop.

GreenAgent is a lightweight, reproducible agent that reports latency, energy, and carbon metrics during execution. Phase 1 focuses on correctness and benchmarking reproducibility.