AB
AB
AiBoss

All Categories

project

ClawWork - HKUDS open-source AI Agent economic survival benchmark framework

ClawWork is an open-source AI Agent economic survival benchmark framework developed by the Data Science Laboratory (HKUDS) at the University of Hong Kong, evaluating the earning potential of large models in real-world business environments. The system sets a minimum threshold of $10 for each Agent...

What is ClawWork?

ClawWork is an open-source AI Agent economic survival benchmark framework developed by the Data Science Laboratory (HKUDS) at the University of Hong Kong, evaluating the "earning ability" of large models in real-world business environments. The system sets a $10 starting capital for each Agent, requiring them to earn income and pay token costs by completing 220 real-world professional tasks (covering 44 industries including finance, healthcare, and manufacturing). Running out of capital results in "bankruptcy." Task rewards are calculated based on quality scores, ranging from $82 to $5,000, forcing Agents to make strategic trade-offs between "earning money immediately" and "investing in learning and improvement." ClawWork supports multi-model competition, including GPT-4o, Claude, and Kimi, and provides a React real-time dashboard to monitor survival status, offering the first real-world stress testing environment based on the "use and disuse" principle for AI labor economics research.

ClawWork's main functions

  • 220 real-world career tasksThe dataset comes from the OpenAI GDPVal dataset, covering 44 economic sectors (manufacturing, finance, healthcare, government, retail, etc.), with each task corresponding to a real-world work scenario.
  • Economic responsibility systemEach token requires a fee, and income comes only from completing tasks. Quality is evaluated by GPT-5.2 (0-1 points), simulating cost control pressures in a real business environment.
  • Work vs. StudyAgents need to make strategic decisions every day – whether to work and earn money immediately to survive, or to invest in learning to improve future performance.
  • Multi-model ArenaIt supports mainstream models such as GPT-4o, Claude, GLM, Kimi, and Qwen to compete on the same stage and compare their economic survivability.
  • Real-time dashboardA React-based visual monitoring system that tracks balance, revenue, costs, and survival status in real time.
  • Nanobot/OpenClaw integrationClawMode transforms any Nanobot gateway into an economically tracking AI colleague, supporting both on-premises and cloud deployments.

ClawWork's project address

  • GitHub repositoryhttps://github.com/HKUDS/ClawWork

How to use ClawWork

  • Install dependenciesAfter cloning the GitHub repository, via pip install -r requirements.txt Install Python dependencies and configure the OpenAI API key for task quality assessment.
  • Configure AgentIn the configuration file, set the model type (GPT-4o/Claude/GLM/Kimi/Qwen, etc.), initial funding (default $10), and daily decision-making strategy (work priority or study priority).
  • Load task setImport 220 real-world occupational tasks from the OpenAI GDPVal dataset, or add custom-designed industry-specific tasks.
  • Start the simulationRun the main program to start the economic survival simulation. The Agent will automatically receive tasks, execute them, and obtain quality scores and corresponding rewards.
  • Monitoring statusView the Agent's current balance, cumulative revenue, token cost consumption, and lifespan through the React real-time dashboard.
  • Analysis resultsAfter the simulation, export a complete report, including key indicators such as profitability, task completion quality, and cost efficiency, for model comparison and strategy optimization.

ClawWork Application Scenarios

  • Large-scale economic capacity assessmentThis study compares the earning power and cost control capabilities of different models such as GPT-4o, Claude, and Kimi under real-world business pressures, providing data support for model selection.
  • Agent Business Feasibility Verification: To test whether the AI Agent has the economic viability to sustain itself before being deployed in actual business, thereby reducing deployment risks.
  • AI Labor Market ResearchIt provides an experimental environment for economists to study the behavior patterns, pricing strategies, and survival rules of AI agents in simulated labor markets.
  • Agent strategy optimizationBy adjusting the "work vs. study" decision parameters, the optimal resource allocation strategy can be found to improve the long-term profitability of agents.
  • Enterprise cost-benefit analysisIt helps companies assess the true cost-benefit of introducing AI agents to replace human labor and calculate the return on investment.
  • Education and training scenariosIt is used to train the next generation of AI engineers to understand the design and business logic of the Agent economic system and to cultivate a cost-conscious development mindset.
Back