Dr. Claw: An AI Scientist Workspace for Vibe Research

Dr. Claw. An AI Scientist Workspace for Vibe Research. By Dingjie Song and twelve others, from Lehigh University, the University of Illinois Chicago, and four other institutions. Published on the thirty-first of August. Its code repository is worked out: about four hundred and fifty source files, with a readme and a licence.

Command-line coding agents can already read and write files and hold a long session together. Claude Code and Gemini CLI are the examples given. What they cannot do is keep a whole research project in one place. End-to-end research still fragments across chat tools, IDEs, terminals and writing environments, and as it fragments, the decisions that would make it auditable are rarely preserved anywhere at all.

Dr. Claw is an open-source workspace built around a deliberate refusal. It does not introduce another autonomous agent. It wraps the coding-agent executors that already exist inside a workflow that stays controllable and auditable, with a human in the loop. The bet is that what has been missing is orchestration rather than more autonomy.

Three components carry that. Persistent state objects hold what has been decided. A reusable skill library holds what has been learned how to do. Multi-executor coordination links the human decisions to the execution that follows them. Together they turn planning, execution and writing into a single loop that is traceable and recoverable, instead of three environments a person copies between and a trail nobody kept.

The workspace is demonstrated through an interactive three-view scenario and a failure-recovery walkthrough, then evaluated against a bare command-line agent sharing the same backend executor. That control is the whole point of the comparison. Holding the executor fixed means what is being contrasted is the orchestration layer, the task graph and state objects and skill library, against the agent it wraps. Under that comparison Dr. Claw scores higher on research completeness while persisting an auditable, recoverable process trail.

The repository is released under AGPL-3.0, with GPL-3.0 upstream components. What the design argues is that auditability is not something added to an autonomous agent afterwards. It comes from holding the human decisions and the machine execution in one recorded structure from the start.

Dr. Claw. An AI Scientist Workspace for Vibe Research. By Dingjie Song and twelve others, from Lehigh University, the University of Illinois Chicago, and four other institutions. Published on the thirty-first of August. This is the long version: the whole paper rather than the abstract.

Command-line coding agents can already read and write project files and sustain context across long sessions. Claude Code and Gemini CLI are the named examples. What they optimize, the paper argues, is execution and not control. The plan, the intermediate decisions, and the artifacts that would make a research process reviewable end up scattered or lost, and the human has few explicit points at which to take over. The diagnosis that follows is that the bottleneck has moved. It is no longer isolated capability. It is full-process orchestration: researchers still switch between tools for decomposition, scheduling, tracking, validation and writing, and every switch weakens reproducibility and delivery reliability.

The paradigm proposed is called Vibe Research, and its definition is more careful than the name suggests. A researcher states high-level goals and constraints in natural language. The system compiles them into an executable research loop and carries out the five core operations: literature review, idea generation, code implementation, results analysis, and drafting. Final acceptance rests on observable outputs. What makes it a paradigm rather than a product is the division of labour it fixes. AI handles the work that is high-throughput, parallelizable and templatable. Humans own research direction, evaluation criteria, key trade-offs and final acceptance. Autonomy is not the axis being argued over. Who decides is.

Four persistent state objects carry the design: a Task Graph, an Artifact Store, a Decision Log, and an Execution Trace. Together they convert interaction history into a reviewable process. A task interaction is modelled as a state transition, where the state at each step is those four objects taken jointly, updated by an action and by the feedback observed. The stated point of that formulation is that the system optimizes iterative state updates rather than single responses. Alongside them sits a skill library. A skill is a directory with a manifest declaring a name and a description, and commonly a version, a license, allowed tools and stage tags. Skills are versioned and schema-checked before activation, and reach a task by three routes: a stage map suggests them, keyword detection over the task text auto-loads them, or a user invokes one by hand. Fifty-eight are mapped across the five research stages, with a hundred and seventy-one in the deployed catalogue. Permissions are treated as a first-class mechanism, with an action allowed only if the current policy admits it.

The evaluation is small, and is reported as small. Holding the executor fixed, Dr. Claw is compared against the bare agent on the three research-hygiene elements where the two conditions diverge. The bare agent produced a limitations section on a third of the tasks; Dr. Claw produced one every time. Related-work coverage moved the same way, from a third to all. The starkest gap is citations, where the bare agent produced zero across all three tasks and Dr. Claw reached two-thirds. The single tie is reported as well, with its reason: on the clinical-note task Dr. Claw's own reference audit did not fire, so it missed citations too.

Two limits are stated with unusual candour. The first is that evaluation: one run per task, three tasks, no statistical power, complemented by a retrospective human study rather than a controlled one. The second is a comparison the authors could easily have left out. In their own design-space table they record that ResearStudio allows a pause at any moment during a run rather than only at stage boundaries, and they write that it is stronger than Dr. Claw on exactly that axis. Mid-run takeover at arbitrary points is the capability a controllability paper would most want to claim, and they decline to claim it.

The framing to take away is that this is a paper about where state lives. An autonomous research agent keeps its reasoning inside a session and hands you a result. The bet here is that the reasoning is half the deliverable, and that a process nobody can inspect is not reproducible whatever it produced. Four state objects, versioned skills, a permission policy bounding the executable action space, and a loop you can revise, retry or take over without losing the state you had. It is released under AGPL-3.0, with GPL-3.0 upstream components.