Research

How HKUDS Built ClawWork to Test How Much AI Agents Can Earn

Built by HKUDS Research Lab (University of Hong Kong) · · Source: GitHub

Quick answer

HKUDS built ClawWork, an open-source benchmark where an AI agent starts with $10 and earns pay for completing 220 professional tasks across 44 occupations. Pay is the graded quality score times the estimated hours times the BLS wage for that job.

Result: According to the repo's leaderboard, the top agent (ATIC + Qwen3.5-Plus) grew a $10 starting balance to $19,915.68 while spending $8.70, with 61.6% average quality. The repo banner puts this at $19K in 8 hours.

What they built

HKUDS published ClawWork, an open-source (MIT) project the repo titles ClawWork: OpenClaw as Your AI Coworker. It describes itself as an economic benchmark: an AI agent starts with $10, does professional work, and earns income based on how good that work is. The repo says it is built on HKUDS's own Nanobot framework and is meant for educational and research use.

The dollar figures are the benchmark's own calculation. Each task pays the quality score multiplied by the estimated hours and the BLS hourly wage for that occupation.

How it works

  • The tasks: 220 tasks from OpenAI's GDPVal dataset, covering 44 occupations. Deliverables include Word documents, spreadsheets, PDFs, project plans and research reports. According to the repo, task values run from $82.78 to $5,004.00, averaging $259.45.
  • The costs: token costs are deducted from the agent's balance after every LLM call, and web searches cost money too. The agent has to earn more than it spends to stay solvent.
  • Work or learn: each day the agent chooses between doing a task and learning, which saves notes to a persistent memory.
  • The tools: eight in standalone mode, including web search, file creation (.txt, .xlsx, .docx, .pdf), Python in a sandbox (E2B by default), a status check for balance and costs, and work submission.
  • The grading: GPT-5.2 scores each submission from 0.0 to 1.0 using category-specific rubrics.
  • The dashboard: a React frontend on a FastAPI backend shows balance, income, costs, quality scores and what the agent learned.

There is also a ClawMode integration that adds the four economic tools to a regular Nanobot agent and puts a cost footer on each response. Setup is a git clone, a Python 3.10 environment, an npm install for the frontend, and API keys in a .env file. An OpenAI key is required.

The result

The repo's leaderboard lists seven agents, all starting at $10:

  • ATIC + Qwen3.5-Plus: $19,915.68 balance, $8.70 in costs, $2,285.31 per hour, 61.6% average quality.
  • Gemini 3.1 Pro Preview: $15,661.71 balance, $105.76 in costs, 43.3% average quality.
  • Qwen3.5-Plus: $15,268.13 balance, $6.78 in costs, 41.6% average quality.
  • ATIC-DEEPSEEK: $10,877.01 balance and the highest average quality on the board at 66.8%.

The banner reads $19K in 8 Hours, while the repo description says $15K earned in 11 hours. HKUDS says top agents reach the equivalent of more than $1,500 per hour. Note that the best average quality score is about two thirds, so a lot of the submitted work was graded well below full marks.

How to build something similar

  • Pick a small set of real tasks from your own work with a clear deliverable for each, and put a dollar value on them based on what you would pay a person.
  • Track what the agent spends per task (tokens, searches, sandbox time) so you can compare cost against the value of the output.
  • Grade the output with a written rubric, ideally using a different model than the one doing the work, and spot check the grades yourself.
  • If you want a head start, clone ClawWork and run it with your own model and keys before building your own version.

The original post

HKUDS · View on GitHub

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases