Claw AI Lab: An Autonomous Multi-Agent Research Team

Fan Wu, Cheng Chen, Zhenshan Tan, Taiyu Zhang, Xinzhen Xu, Yanyu Qian, Dingcheng Gao, Lanyun Zhu, Qi Zhu, Yi Tan, Deyi Ji, Guosheng Lin, Tianrun Chen, Deheng Ye, Fayao LiuarXiv 2026 · arXiv

TL;DR

Claw AI Lab shifts autonomous AI research away from black-box, "prompt-to-paper" pipelines toward an interactive, multi-agent laboratory environment. By introducing a structured five-layer architecture and a specialized execution harness, it allows users to monitor, intervene, and reliably execute code and experiments. The system demonstrates significant improvements in experiment completeness and paper quality over existing baselines.

What Problem Does This Paper Address?

Current autonomous research agents typically operate as linear, hidden pipelines. While they can generate surface-level text, they often suffer from critical practical failures: experiments run only partially, intermediate code artifacts are impossible to inspect or debug, and final papers sometimes contain fabricated tables that do not faithfully reflect actual execution logs.

Claw AI Lab addresses this lack of steerability, visibility, and reliability. Instead of treating AI research as a one-shot document generation task, the authors reframe it as an interactive, persistent, and inspectable AI laboratory operation.

Methods

Claw AI Lab adopts a pyramid-style, hierarchical multi-agent architecture organized into five connected layers:

  1. Idea Layer: Uses a multi-agent discussion phase where multiple agents propose parallel ideas, engage in structured debate, and use a consensus mechanism to select the most promising direction.
  2. Planning Layer: Decomposes the chosen idea into concrete tasks, dependencies, and milestones, iteratively refining them through a validation loop.
  3. Coding Layer: Centered on the novel Claw-Code Harness, agents inspect local codebases, datasets, and checkpoints to write and debug code. It executes tasks in a sandboxed workspace and injects a read-only Python controller to enforce time budgets, catch NaN/Inf errors, and run anti-fabrication checks to detect placeholder code or fake metrics.
  4. Experiment Layer: Deploys experiments on computational resources and collects metrics. Crucially, it supports cross-layer feedback—unexpected results or repeated failures can trigger updates back in the planning or idea layers.
  5. Writing Layer: Transforms experimental outcomes into structured research outputs by generating outlines, producing visualizations, and drafting the manuscript.
Figure 1: Overview of Claw AI Lab. The system organizes automatic research into five connected layers: idea, planning, coding, experimentation, and writing layers. Each layer uses specialized agents and validation loops, while feedback can flow across layers to revise earlier decisions when needed.

Datasets

The platform was evaluated internally across four diverse AI research and reproduction case studies rather than a single static dataset:

  • Topic 1: "Quantifying Hallucination in Generated Video Models" (Research)
  • Topic 2: "LIAR Dataset-Based Fake News Classification Solution" (Research, utilizing the LIAR dataset)
  • Topic 3: "A Q-Learning Approach for Student Performance Improvement Using Public Educational Data" (Research)
  • Topic 4: "Reproducing and Analyzing PhyCustom on Flux" (Reproduction)

Key Insights and Ideas

  • Research is Non-Linear: A key "aha moment" is that autonomous research cannot succeed as a pure one-way pipeline. Implementing cross-layer feedback loops allows downstream coding or experimental failures to dynamically reshape upstream plans and ideas.
  • The "Harness" is Essential for Integrity: Simply prompting an LLM to write a paper leads to weak empirical support. By embedding a dedicated execution harness (Claw-Code Harness) that directly links local codebases and datasets to sandboxed execution environments, the system ensures that the final paper accurately reflects real, uncorrupted experimental data.
  • Infrastructure over Automation: The authors argue that the future of AI science lies in building usable scientific infrastructure (dashboards, real-time event streams, and one-click rollbacks) that allows humans to collaborate with agent teams, rather than trying to completely eliminate the human from the loop.

Key Results

The authors compared Claw AI Lab against the AutoResearchClaw baseline using advanced LLM evaluators (ChatGPT 5.4 Thinking and Gemini 3.1 Pro) across six dimensions: technical depth & reproducibility, structure & section flow, novelty & contributions, clarity & terminology, logical argumentation, and citations & evidence support.

  • Research Topics (1–3): Claw AI Lab achieved substantial and stable improvements, with average score gains ranging from +15.5 to +16.5 points out of 100.
  • Reproduction Topic (4): The system improved the baseline reproduction report score from 73.0 to 78.0, marking a +5.0 point gain.
Figure 2: Detailed comparison for four paper pairs scored by Gemini and ChatGPT, respectively. Each panel compares Claw AI Lab with AutoResearchClaw on six dimensions.

Limitations and Future Work

While Claw AI Lab represents a significant step toward interactive science automation, its performance remains tightly coupled with the baseline reasoning capabilities of its underlying frontier models (such as GPT-5.4 and Gemini-3-Pro) for complex debugging and novel hypothesis generation. The paper positions this work as an early foundation, implying that future iterations will need to focus on even tighter human-agent collaboration and scaling up to more complex, multi-project engineering environments.

Summary updated

Abstract

We present Claw AI Lab, a lab-native autonomous research platform that advances automated research from a hidden prompt-to-paper pipeline into an interactive AI laboratory. Rather than centering the system around a single agent or a fixed serial workflow, we allow users to instantiate a full research team from one prompt, with customizable roles, collaborative workflows, real-time monitoring, artifact inspection, and rollback/resume control through a unified dashboard. The platform also supports distinct research modes for exploration, multi-agent discussion, and reproduction, making autonomous research substantially more steerable and laboratory-like in practice. A key practical contribution of Claw AI Lab lies in its Claw-Code Harness, which connects local codebases, datasets, and checkpoints to runnable experiments and feeds execution artifacts back into the research loop. As a result, the harness improves not only execution integration, but also experimental completion and result integrity: experiments are easier to inspect, iterate on, and faithfully transfer into final papers, reducing common failure modes such as partial runs and malformed result reporting. In our internal evaluation on five AI research case studies, using AutoResearchClaw as the baseline, Claw AI Lab is consistently preferred by AI expert judges on idea novelty, experiment completeness, and paper presentation quality. We view Claw AI Lab as an early step toward a new paradigm: autonomous research as usable, interactive, and reliability-aware scientific infrastructure.
BibTeX
@misc{wu2026claw,
  author = {Fan Wu and Cheng Chen and Zhenshan Tan and Taiyu Zhang and Xinzhen Xu and Yanyu Qian and Dingcheng Gao and Lanyun Zhu and Qi Zhu and Yi Tan and Deyi Ji and Guosheng Lin and Tianrun Chen and Deheng Ye and Fayao Liu},
  title = {{Claw AI Lab: An Autonomous Multi-Agent Research Team}},
  year = {2026},
  eprint = {2605.22662},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.22662},
}

Loading related content...

Figure