ApexClaw: A Persistent Infrastructure for Agentic Scientific Research

Most AI-for-science systems (agents4science) are evaluated as task-specific automation rather than persistent work environments. This leaves an important and pragmatic systems-question unresolved: what infrastructure enables scientific agents to be trustworthy, steerable, and reproducible? We present ApexClaw, a persistent workspace for human-supervised AI agents in scientific discovery. The system provisions isolated Linux environments with scientific computing libraries, a browser-based IDE, and a registry of reusable scientific skills. Rather than pursuing full autonomy, ApexClaw emphasizes human oversight through an interface that exposes agent reasoning, tool calls, and workspace files, allowing scientists to guide agents at key junctures. We validate this approach with telemetry from 81 users operating 243 workspaces across 252 conversations (March--May 2026): 87\% of users reconnected to existing workspaces, and humans intervened selectively on 3.4\% of interactions. These findings demonstrate that hybrid autonomy with persistent context is ordinary practice in real agentic science, and that durable workspaces, explicit artifacts, and auditable traces are necessary infrastructure for reproducible and steerable AI scientists.