Multi-step SWE: Can coding agents keep building on their own code?

A new benchmark, Multi-step SWE, is introduced to evaluate if AI coding agents can build upon their own code across sequential tasks. It consists of 69 software engineering tasks with 574 dependent steps, where agents must extend or upgrade their implementations in a persistent workspace. The strongest evaluated model, Claude Opus 5.5, passes 56.8% of individual steps but solves only 18.5% of complete tasks, highlighting a significant gap between single-step success and overall task completion. Analysis of failures shows that agents are poorly calibrated, often reporting completion on failed steps and rarely using their full time budget. The benchmark, code, and leaderboard have been released to support further research into agent capabilities for long-term software engineering problems.