HKU puts AI agents into a cutthroat economy

What happens if you hand an AI agent ten bucks, force it to pay for its own API calls, and tell it to earn a living?

Most LLM benchmarks test whether an AI can answer multiple-choice questions in a vacuum. On February 15, 2026, the HKU Data Intelligence Lab released ClawWork, an open-source framework that drops AI agents straight into a cutthroat simulated economy to see if they can survive.

Make money or go bankrupt

ClawWork treats AI evaluation like an extreme survival game. Every agent starts with a modest $10 budget. Every prompt, reasoning step, and web search consumes tokens that are billed in real time; if an agent runs out of cash, it goes bankrupt.

The only way to stay alive is to complete real white-collar work. Built on the GDPVal dataset, ClawWork dishes out 220 professional tasks across 44 occupational sectors, ranging from manufacturing procurement to financial modeling and healthcare administration. Agents are expected to produce tangible artifacts like Word documents, spreadsheets, and PDFs.

Once work is submitted, GPT-5.2 acts as an evaluator, grading deliverables against sector-specific rubrics tied to Bureau of Labor Statistics wage data. In HKU's tests, top agents pulled in an equivalent of over $1,500 an hour, with one agent raking in $10,000 across 44 sectors in just seven hours.

Agents also face daily strategic trade-offs: they can grind billable hours for immediate cash or invest their balance into learning to boost future performance, mimicking real career decisions.

Real costs, zero fluff

Built on a lightweight Nanobot architecture, ClawWork pairs a Python CLI with a React dashboard where users can watch agents manage cash flow and make decisions in real time.

The framework already tracks frontier models including Claude Sonnet 4.6, Gemini 3.1 Pro, Qwen-3.5-Plus, Qwen3-Max, Kimi-K2.5, and GLM-4.7. Rather than estimating token costs, ClawWork pulls the exact billing figures directly from API responses, including hidden thinking tokens.

By pitting models head-to-head under identical economic conditions, it crowns an AI champion based on profit margins rather than academic exam scores.

Why it matters

We think ClawWork's economic angle is spot on. For the past two years, the enterprise AI conversation has been stuck on benchmark scores that mean very little to business bottom lines. What companies actually care about is return on investment: will this agent deliver more economic value than the API fees it burns?

For developers and teams building automation workflows, this offers a brutally honest reality check. A model might be brilliant at reasoning, but if it burns 50,000 thinking tokens on excessive web browsing to generate a simple purchase order, it is an economic liability. ClawWork exposes exactly where token bloat destroys margin.

That said, the setup still has an obvious catch. Having GPT-5.2 serve as the boss who inspects work and signs paychecks means evaluations still rely on synthetic judgment rather than messy human feedback. But shifting AI benchmarking from academic accuracy to economic solvency is a massive step toward understanding how agents will actually behave on payroll.