Salesforce's DarwinX Framework Boosts Self-Improving AI Agents by Up to 49.5%
Salesforce researchers have made significant breakthroughs in developing self-improving AI agents. Their new DarwinX framework takes an evolutionary approach to improving performance without relying on manual rewriting of agent harnesses.
The team addressed two major problems with existing self-improvement loops: 'path dependence' and 'cross-task interference.' Path dependence occurs when early changes determine the foundation for future evolution, potentially leading to plateaus. Cross-task interference happens when a change helps one class of tasks but damages another.
DarwinX treats these issues as a selection problem. It generates different variants of the harness, stores them in an archive, and only lets improvements advance if they improve capability without regressing on other tasks.
The researchers tested DarwinX on four benchmarks, achieving significant gains: 3.4-point improvement on SWE-bench Verified to 49.5-point jump on WebArena-Infinity.