RSIHub: A Composable Framework for Unified Evaluation of Agent-Evolution Methods

01 ABSTRACT

Agent-evolution methods use execution feedback to improve prompts, skills, tools, and orchestration, but implementations often rebuild the same experimental machinery. RSIHub is an open-source framework that composes methods via reusable operators and mutable resource declarations, supplying execution, evaluation, recovery, and lineage. The authors instantiate four methods with MiniSWE and Codex on Terminal Bench 2 and τ^3 Banking; the fact is a largest observed overall pass@1 gain of 13.5 percentage points, and the authors' view is that this provides a common protocol for comparing evolution policies.

02 KEY FINDINGS

1. RSIHub composes reusable operators in recipes and declares mutable resources, keeping method-specific logic in recipes and operators so AHE, Hyperagents, A-Evolve, and GEPA can be instantiated without modifying the framework core.
2. The framework supplies execution, evaluation, recovery, and lineage infrastructure; evaluation stays outside the mutable surface, and results remain tied to candidates and evaluation conditions.
3. Under fixed model and evaluator configurations, evaluated with MiniSWE and Codex on Terminal Bench 2 and τ^3 Banking, the largest observed overall pass@1 gain is 13.5 percentage points.