SWE-agent: Interfaces Change Outcomes
SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
Episode participants
A solo episode without invited guests.
What we will explore
Live on October 5, 2026 at 12:00 MSK (09:00 UTC). In this solo episode, Alexander Polomodov reviews the NeurIPS 2024 SWE-agent paper: how the interface between an agent and its computer changes what the same language model achieves in a repository.
We will examine code search, file views with line numbers, multiline edits, syntax checks and context selection. We will trace how actions and environment feedback help an agent stay oriented, verify changes and recover from mistakes.
We will compare SWE-bench and SWE-bench Lite, settings with and without demonstrations, ablations and trajectories. We will distinguish historical 2024 measurements from lessons for current models: transfer the method for testing an interface, rather than numerical settings from one experiment.