Retrievolver: A Self-Evolving Harness with Dynamic Workflows for Heterogeneous Retrieval

Agents need evidence from text, code, images, document pages, and the web, yet each source typically needs its own retriever, often trained in-domain, or a hand-built pipeline. We present Retrievolver, a self-evolving retrieval harness with a common interface for tasks, tools, and evidence. A planner with fixed model weights writes a workflow for each query. The workflow combines specialist retrievers and rerankers. The planner follows a Policy and Skills stored as readable documents. Users provide a corpus and tools, a few hundred training and development queries, and a quality evaluator. Retrievolver adds task-relevant methods from papers and author code to its tool pool, then revises its guidance by tracing where evidence was lost during search, ranking, or delivery. Development evaluation selects the version frozen for testing. In a five-task text ablation with a shared tool pool and fixed model weights, evolved guidance raises macro nDCG@10 by 8.8 points over generic Initial guidance and 7.5 over a fixed pipeline. Across 42,721 test queries in five task families, Retrievolver leads four, and its development-selected systems, which also expand tools and adapt their execution limits, improve the primary metric of all five over their initial systems. Transfer to five target benchmarks uses no target relevance labels, and the frozen retrieval module raises BrowseComp-Plus answer accuracy in all five agent hosts.