Independent · Reproducible · Pre-launch
Do autonomous AI agents actually get the job done?
We run popular agents through the same real tasks, three times each, and score them on finish rate, cost, and how often a human had to step in.
Testing is in progress. Nothing on this site is a result yet; scores appear only after real runs are completed and logged.
Same tasks
Five standard tasks, identical prompts, fresh state, three runs per agent.
Hard numbers
Finish rate, median cost per run, average human interventions, and a 100-point score.
Open logs
We plan to publish prompts, versions, dates, and raw logs so anyone can check our work.
Agents on the test list
Shortlist of 10, pending verification of each product's current plan and access.
Claude CodeCodexDevinGoogle AntigravityCursor agentopencodeManusChatGPT WorkClaude CoworkGenspark