Introducing Sonar Foundation Agent

Sonar Foundation Agent, developed by the former AutoCodeRover team at Sonar, is a coding agent designed to address general software issues. By November 2025, it achieved a 75% score on SWE-bench Verified, with an efficiency of 10.5 minutes per issue and an average cost of $1.26. Implemented using the LlamaIndex framework, Sonar Foundation Agent utilizes a tool-calling approach to iteratively solve coding problems, ultimately producing a unified diff format patch. The agent is equipped with tools like bash for executing commands, str_replace_editor for file editing through string replacement, and find_symbols for searching program ASTs. The evolution from AutoCodeRover to Sonar Foundation Agent involved moving from a constrained two-stage workflow to a more autonomous single-agent process, thanks to advances in large language models (LLMs) like GPT-5 and Claude Sonnet 4.5. This shift allowed the agent to handle longer context windows and follow instructions more effectively, boosting its efficacy from 58% to 75%. The team also refined prompts to be less prescriptive, enhancing the agent's performance by capitalizing on the thinking models of advanced LLMs. This progression underscores the importance of aligning agent autonomy with the capabilities of underlying models in AI-driven software development.