SWE-Together: Grade Coding Agents on Multi-Turn Sessions

Why SWE-bench scores lie about coding agents

On July 5, Meta open-sourced SWE-Together, a 109-task benchmark that scores a coding agent across an entire interactive session instead of a single prompt. The first headline result: Claude Opus 4.8 needed the least corrective steering of any model tested, finishing 63% of multi-turn tasks (pass@1) with the fewest human interruptions. The number that matters there is not 63%, it's "least steering" — SWE-Together measures how much you have to babysit the agent, not whether it can one-shot a clean GitHub issue.