Dr. Claw: An AI Scientist Workspace for Vibe Research

Dingjie Song, Hanrong Zhang, Dawei Liu, Yixin Liu, Zongxia Li, Zhengqing Yuan, Siqi Zhang, Henry Peng Zou, Zhiling Yan, Yuxuan Zhang, Yanfang Ye, Philip S. Yu, Lichao Sun

Dr. Claw unifies AI research into a controllable, auditable workspace by wrapping existing coding agents in a persistent task-graph orchestration layer.

How can we unify the fragmented research workflow—problem definition, experimentation, and writing—into a single, state-aware AI workspace?

Research agents can execute code, but they lack the structure to manage long-horizon projects, leaving researchers to manually track progress across fragmented tools and lost decision histories. Dr. Claw wraps these agents in a persistent orchestration layer that maps high-level goals into a structured task graph, maintaining an auditable log of every decision, artifact, and execution step. In open-ended research tasks, this integration improves completeness by reliably enforcing research hygiene—such as limitations sections and literature citations—that bare agents frequently omit.

Paper Primer

Dr. Claw models research as a state-transition loop where the system manages four persistent objects: a Task Graph, Artifact Store, Decision Log, and Execution Trace. It functions as a wrapper: it delegates high-throughput execution to existing command-line coding agents while the orchestration layer handles task decomposition, skill-based tool selection, and human-in-the-loop checkpoints.

Dr. Claw increases research completeness in open-ended tasks by enforcing best-practice hygiene.

In a three-task medical research pilot, Dr. Claw achieved a pooled completeness score of 0.952 compared to 0.873 for the bare agent, specifically closing gaps in subgroup analysis and literature citations.

The system provides a reusable library of 171 skills mapped to five research stages, allowing users to trigger specific capabilities (e.g., "reference audit" or "subgroup analysis") that the underlying agent might otherwise ignore. Because the orchestration layer persists the entire project state, users can recover from failures by revising or retrying specific tasks without wiping the existing artifact trail.

Why wrap an existing agent instead of building a new, specialized research executor?

The authors argue that the bottleneck in AI research is not isolated execution capability, but process orchestration and context maintenance. By wrapping existing agents, Dr. Claw reuses proven coding performance while adding the auditability and human-control affordances that bare executors lack.

Does this system automate the research process entirely?

No; Dr. Claw is designed for "Vibe Research," a human-in-the-loop paradigm where the AI handles templatable execution while the human retains ownership of research direction, evaluation criteria, and final acceptance.

The Research Fragmentation Problem

We expose why fragmented tools hinder reproducible research and introduce a unified workflow.

Current research pipelines are split among disparate tools—chat interfaces, IDEs, terminals, and document editors—so the plan, intermediate decisions, and artifacts become scattered and hard to audit.

When a researcher hops between separate applications, each step lives in its own silo, making the overall workflow opaque and error‑prone.

The shift from tool‑switching to continuous research workflows eliminates hidden hand‑offs and preserves a complete, auditable process.

Existing Research Automation

We position Dr. Claw among prior agents, highlighting its CLI integration, persistent state, and full‑workflow control.

Prior work splits into three strands: end‑to‑end automation, low‑code authoring, and intervenable agents. Dr. Claw unifies these strands by wrapping a CLI coding agent, persisting four research‑state objects, and enabling continuous human takeover.

A CLI coding agent is a program that receives textual commands on a terminal and emits source‑code files as its response.

We evaluate Dr. Claw’s impact on workflow efficiency, quality, and usability rather than on raw model performance.

The Dr. Claw Architecture

Dr. Claw turns research into a single, continuously orchestrated workflow.

Researchers currently hop between separate tools for framing a problem, running experiments, and writing papers. Dr. Claw’s answer is to keep the entire process inside a single, state‑aware loop, so every decision and artifact is visible and revisitable.

The four objects form the immutable backbone of a research project, recording what to do, what has been produced, what decisions were made, and how the system progressed.

Step 1: User adds a constraint to T1, creating a new node in the Task Graph.

Step 2: AI executes T1, producing artifact v1; the Artifact Store records $v1$ and links it to T1.

Step 3: The Execution Trace logs “T1 completed, success=true”.

Step 4: Decision Log records the human’s acceptance of $v1$ and schedules T2.

Step 5: State $S_{t+1}$ now reflects the updated Task Graph, Artifact Store (v1), Decision Log (acceptance), and Execution Trace (completion).

Even this two‑task toy shows how every modification flows through the four objects, making the whole workflow reversible and inspectable.

Interaction layer: the researcher writes a high‑level goal and uploads initial artifacts.

Orchestration layer: the goal is parsed into a Task Graph node, a suitable skill is attached, and the Decision Log records the intent.

Execution layer: the attached skill runs, updates the Artifact Store with new results, and appends an entry to the Execution Trace.

Orchestration checks the safety policy; if the action is disallowed, the system pauses and prompts the user.

After verification, the Decision Log records acceptance, and the Interaction layer presents the updated state to the researcher.

Permission check before any external call.

**Figure 1.** Control split in Dr. Claw: the Human-Controlled side sets goals, constraints, and acceptance decisions; the AI-Execution side runs the five core AI operations (Plan $arrow$ Execute), connected through four state objects and checkpoint feedback (Verify/Revise/Retry/Handoff).

**Figure 2.** **Overall system view of Dr. Claw.** Top: four state objects (Task Graph, Artifact Store, Decision Log, Execution Trace). Middle: three system layers (Interaction, Orchestration, Execution). Bottom: iterative workflow steps (Plan–Execute–Verify–Write-back). Lowest: safety policy constraint ($Action_t \in \mathcal{A}(Policy_t)$).

The four state objects serve as the single source of truth for research continuity, ensuring that every artifact, decision, and execution step is persistently recorded and instantly accessible.

How does Dr. Claw’s state‑object model differ from a simple relational database that stores tasks and results?

A relational database is passive storage; Dr. Claw’s objects are active participants in a feedback loop. The transition function $f$ continuously rewrites the objects based on actions and observations, enabling reversible, versioned, and policy‑aware updates rather than static rows.

Human-in-the-Loop Research Workflow

Illustrates how Dr. Claw turns a high‑level research intent into a controllable, auditable workflow.

Research tasks often stall because high‑level intent cannot be reliably turned into concrete, repeatable steps. The core trick is to keep the entire process inside a persistent, state‑aware workflow so that every decision is visible and reversible.

Human oversight is kept at the decision points while the system automates decomposition, execution, and bookkeeping.

Dr. Claw parses the prompt and creates a Task Graph with two nodes: (1) literature review, (2) experiment setup.

The human reviews the generated brief and approves node 1.

Node 1 is dispatched to a literature‑search executor; results are written to the Artifact Store.

The Decision Log records the approval and the executor’s output.

Node 2 is then presented; the human tweaks the hyper‑parameter range before confirming.

Execution proceeds, the final model artefact is stored, and the Execution Trace logs the full run.

This concrete loop shows how the system preserves human control while automating the heavy lifting, and how every artefact is traceable for later recovery.

Generate a structured brief from the high‑level research prompt.

Construct a dependency‑aware Task Graph and schedule tasks.

Execute each task (Plan → Execute) on heterogeneous backends.

Verify outputs against the acceptance criteria.

Write results into the Artifact Store and update the Task Graph.

Record every human decision and system event in the Decision Log and Execution Trace.

The centre panel of Figure 3 is the primary orchestration view. A single unified research prompt drives the whole process, while the user (research lead) confirms goals and key decisions. Dr. Claw handles decomposition, dispatch, and state feedback, and the four state objects stay synchronized throughout.

**Figure 3.** Three-view scenario diagram. Left: Skills Dashboard for capability discovery and filtering. Center: main orchestration interface for planning, execution feedback, and human approval. Right: Task List for progress tracking and next-step selection.

The left panel demonstrates capability management. Users browse the Skills Dashboard by research stage, filter by tags (e.g., Ideation, Experiment), and can manually add new skills to the current project for later invocation.

The right panel shows the Task List interface. It synchronises task status, displays a progress bar, and groups tasks by stage. Users can inspect progress, trace each task’s ID and linked skill tags, or trigger the next pending step directly.

The core trick is a single, persistent research prompt combined with unified state objects, which lets Dr. Claw swap backends, preserve audit trails, and keep the human in control.

Performance and Recovery Evaluation

Dr. Claw improves research completeness and auditability while preserving output quality.

Recall that Dr. Claw models research as a persistent, state‑aware workflow of four core objects. This section shows how that design impacts completeness and auditability.

Dr. Claw achieves higher pooled research completeness than the bare command‑line agent.

Across three medical tasks, Dr. Claw attained a pooled completion of 0.952 versus 0.873 for the bare agent.

**Figure 4.** Open-ended completion per task (fraction of 21 elements); dotted lines are pooled means.

**Figure 5.** Pooled pass rate on the three research-hygiene elements where the conditions diverge.

**Figure 6.** Whether a run persists a queryable process model (present/absent by design, not scored).

**Figure 7.** Captured failure and non-destructive recovery inside a project. Left: the induced wrong-path error, captured in the execution trace. Right: the in-place fix and recovered result, with all prior files retained.

Ethics and Limitations

We acknowledge funding sources and outline the ethical safeguards built into Dr. Claw.

We thank the National Science Foundation for partial support via grants CRII‑2246067, ATD‑2427915, POSE‑2346158, and POSE‑2449280.

Dr. Claw is designed to assist research under sustained human control and never operates autonomously. Every stage of generation passes through explicit human checkpoints, and the Decision Log and Execution Trace record all approvals, revisions, and rejections.

Critical content such as citations, experimental conclusions, and manuscript claims must be verified by a researcher before use. When handling sensitive data, users should apply least‑privilege permission settings; in our study the medical datasets remained on the authors’ server and no patient‑level data were released. For high‑risk domains like healthcare, system outputs must not be employed directly for real‑world clinical decisions, and the human study participants provided informed consent while AI‑based coding assistants were used in accordance with venue policy.

Human Study and Retrospective Analysis

Ablation results show that removing Dr. Claw’s components degrades time, performance, and experience across research stages.

Recall that Dr. Claw unifies fragmented research tools by modeling the workflow as a persistent, state‑aware process over four core objects.

Dr. Claw yields significantly shorter completion‑time bands than both No‑AI and Web/Desktop‑AI across all stages.

Friedman $\chi$²(2)=13.56, p=0.0011; pairwise Wilcoxon p = 0.0469 for both controls.

Stage‑output performance scores are highest for Dr. Claw, intermediate for Web/Desktop‑AI, and lowest for No‑AI.

Friedman $\chi$²(2)=9.25, p=0.0098; pairwise Wilcoxon p = 0.0469 (Dr. Claw vs No‑AI) and p = 0.0625 (Dr. Claw vs Web/Desktop‑AI).

Experience scores are strongest for Dr. Claw, with pairwise significance against both baselines.

Friedman $\chi$²(2)=12.29, p=0.0021; Holm‑corrected pairwise p = 0.0469 for all three stages.

**Figure c.** Mean performance (Perf) and experience (Exp); error bars: SE.

This heatmap displays the proportion of switch-count bands across three stages (Stage 1, Stage 2, Stage 3) for three different systems (N=No-AI, W=Web/Desktop-AI, D=Dr. Claw). The color intensity represents the proportion, ranging from 0% (white) to 100% (dark red).

**Figure a.** Time-band distributions (h/d/w/m: hour/day/week/month).

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers