Research
Combining complementary coding agents with DEI
A committee reviews candidate fixes produced by different software-engineering agents and selects a patch for a repository issue.
Software-engineering agents can resolve different sets of repository issues even when their overall scores look similar. Diversity Empowered Intelligence, or DEI, studies how to use that complementary behavior through a committee of agents.
Collect several candidate fixes
DEI operates as a layer above existing software-engineering agent frameworks. Those agents attempt an issue and produce candidate patches. The framework then brings the candidates into a common review process.
This construction preserves the individual agent implementations. Their different approaches to exploring a repository and proposing a change become a source of variation in the patches available for selection.
Make review a separate component
The committee compares candidate fixes and uses a reranking process to select an answer. The paper evaluates groups formed from different agents as well as multiple runs of an individual agent.
Its experiments on SWE-Bench Lite examine both the diversity of resolved issues and the performance of the selection mechanism. The results connect the benefit of a group to the available candidates and the committee's ability to choose among them.
Study the complete ensemble
The work contributes an orchestration approach for integrating existing coding agents and an experimental analysis of their complementary strengths. The released implementation provides a way to inspect how candidates enter the selection process.
For teams studying agent ensembles, the framework suggests two concrete objects to examine. The candidate set shows which solutions the agents produced. The selection stage shows how the system used that set. Recording both makes it possible to investigate whether a change in group composition or review behavior affected the final patch.
Kexun Zhang and collaborators conducted the research at Salesforce AI Research and Carnegie Mellon University. Weiran Yao is a co-author. The work appeared at ICLR 2025. Read the paper and explore the committee implementation.