Sonar’s Foundation Agent Tops SWE-bench, Signaling Step-Change in Autonomous Code Remediation

New updates have been reported about Sonar.

Sonar has placed its new Foundation Agent at the center of its AI strategy for code verification, after the tool secured the highest score on the unfiltered SWE-bench leaderboard, with a 79.2% success rate on SWE-bench Verified and 52.62% on SWE-bench Full. Built by the team behind AutoCodeRover, the agent resolved issues in about nine minutes on average at roughly $1.90 per fix, indicating that high-accuracy autonomous remediation can be delivered at enterprise scale and materially lower cost than manual intervention.

The Foundation Agent, which is LLM-agnostic but achieved peak performance using Anthropic’s Claude Opus 4.5, uses advanced tool-calling via LlamaIndex, stateful bash execution, AST-based symbol search, and test-driven remediation to navigate full codebases and ship patches that pass original project tests. Sonar is using this research architecture to power its SonarQube Remediation Agent (currently in beta) and plans broader integration into its verification platform, positioning itself as the independent control layer for AI-generated code and autonomous agents, even though the Foundation Agent itself is not yet a commercial product.