Research2026-09-03

Qwen released E-Commerce Bench, a test where AI agents run a simulated online shop for a year. Each agent starts with ¥100,000 and opens up to four stores. It must negotiate with suppliers, set prices, hold stock and stay solvent. Nothing resets, and the agent's memory overflows long before day 365. Qwen ran 18 models five times each and scored them on end-of-year assets. Rankings differ sharply by dimension. The top earner sent 18.5% of its buying cash to fraudulent suppliers, against 0.12% for the most careful model. Several models went bankrupt in some runs. Demand and supplier pricing are fixed in advance, so results reflect the agent, not luck. Of 576 suppliers, 152 run one of five scam patterns.

What changed

Agent benchmarks mostly tested short tasks that reset after each attempt.

What it unlocks

Running an open-source year-long business simulation to compare how models handle money, suppliers and memory.

  • ¥100,000 starting balance, 365 days
  • 18 models, 5 episodes each
  • best model ended at ¥1,431k mean
  • 18.5% vs 0.12% spend to scam suppliers

What you need to act on it

  • Python 3.10+
  • API key for the model under test
  • a second key for the supplier role-play model

Sources

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.