Dr. Claw: An AI Scientist Workspace for Vibe Research

Paper: Dr. Claw: An AI Scientist Workspace for Vibe Research (v1) https://arxiv.org/html/2609.00365v1

Dr. Claw: carrying research state through recovery

Recovering a failed research step should not require recreating the experiment’s entire context. Dr. Claw makes persistent plans, outputs, decisions and execution traces the workspace’s organizing structure.

The system wraps existing command-line coding agents rather than introducing a new executor. Its four persistent objects are a Task Graph, Artifact Store, Decision Log and Execution Trace. The graph records tasks and dependencies; the store connects outputs to work; decisions retain interventions; traces preserve actions and observations. A Plan, Execute, Verify and Write-back loop updates these objects while allowing human inspection, revision and takeover.

The original animated dataset-preparation project shows how a changed direction can remain connected to prior work. It is schematic teaching material, not a product screenshot or recorded pilot. Skills supply reusable instructions for project stages, but a suggested skill is different from evidence that it was actually invoked.

The paper’s recovery walkthrough induces a wrong-path failure inside one mini-project. The error enters the trace, execution halts, and recovery proceeds in place with all five prior files retained. Across 23 tool events, corrected artifacts are added without wiping the earlier attempt. This demonstrates a preservation mechanism; it is not a matched comparison of recovery speed or reliability.

A separate pilot uses the same GPT-5.4 backend for the wrapper and bare agent on three medical-research tasks, with one run per task. Completeness measures the presence of 21 research-practice components in produced files. Pooled coverage is 0.952 versus 0.873; the bootstrap interval for the difference includes zero. Coverage is useful to inspect, but it is not scientific correctness. The continuing project remains a record to examine before accepting its conclusions.

Watch the companion English animation with external English, Simplified Chinese or bilingual captions.

Paper, exact arXiv v1: https://arxiv.org/html/2609.00365v1

#ResearchTools #AIAgents #AIResearch