Multi-step SWE: Can coding agents keep building on their own code?
69 tasks, 574 dependent steps, nine frontier models. The strongest model passes 56.8% of steps but solves only 18.5% of complete tasks.
Read the researchStudying how agents learn, build, and act over longer horizons.
69 tasks, 574 dependent steps, nine frontier models. The strongest model passes 56.8% of steps but solves only 18.5% of complete tasks.
Read the research