YouRA (Your Research Agent) is a persistent-state architecture designed for evidence‑traceable autonomous research, addressing the structural gap where end‑to‑end research agents produce manuscript claims that diverge from executed experiments. The key idea is to keep hypotheses, experiment gates, evidence pointers, and failure histories in a verifiable persistent state that spans the entire long‑horizon research pipeline.
The system consists of three main modules. The first is the Verification State Architecture (VSA), which records each hypothesis, its associated gate, and pointers to generated evidence in a structured state graph, enabling continuous alignment between claims and evidence. The second is the Independent Controller, which transforms state and reflection records into lifecycle management, error recovery, and debate/review workflows, decoupling control logic from execution. The third is Stateful Reflection, which logs structured lessons whenever an experiment fails and triggers bounded repair, redesign, or reset pathways based on predefined recovery strategies.
On the predefined ten‑task end‑to‑end subset of MLR‑Bench, YouRA outperforms both MLR‑Agent and AI Scientist V2 on the overall scalar metric, and shows higher proportions of real‑data‑based outputs in four fact‑based hallucination categories and data‑provenance diagnostics. Ablation studies on the VSA, Independent Controller, MCP tool access, and reflection‑guided recovery confirm their separate contributions; removing any core‑state component drops performance below that of the full system.
The implementation is open‑source at https://github.com/PrayPrey/Your-Research-Agent
Review: By introducing explicit state management and a traceable evidence chain, YouRA equips autonomous research agents with robust error recovery and result verification, laying a solid foundation for large‑scale automated science.