CoffeeBenchTrajectories

Benchmarking Long-Horizon LLM Agents in Heterogeneous Multi-Agent Economies

CoffeeBench

CoffeeBench measures how much net income an LLM agent can earn as a coffee roaster over 90 simulated days, inside a multi-agent economy of two farmers, two roasters, and two retailers. The evaluated model runs Roaster A; the other five firms are fixed Claude Sonnet 4.6 baselines.

CoffeeBench: two farmers, two roasters, and two retailers trading on a shared marketplace
Six firms — two at each layer of the supply chain — trade on a shared marketplace, each maximizing its own net income.

Evaluation results

Explore individual runs

Timeseries

Cash

Equity

Net income

Retail price per retailer

Each line is a retailer's average daily price; bubble size is units sold that day.

Consumer-sales revenue per retailer

B2B deal prices over time

Recent deals

day seller buyer item qty unit_price total terms

Messages

time sender recipient title body_preview

Select a row to read the full message body.

Tool calls