We Gave Agents IDE-Native Search Tools. They Got Faster and Cheaper. | The JetBrains AI Blog

The implementation of IDE-native search tools for coding agents significantly improved performance metrics such as latency and cost without compromising quality. By integrating a prebundled search skill with a unified MCP tool, agents could utilize IDE-specific resources like indices and project models, which shell tools like grep and find cannot access. The new system reduced median latency by 8.33% and P95 latency by 16.44%, while total costs decreased by 5.60% and budget overruns fell by 33.28%. The evaluation used paired delta analysis with tasks conducted both with and without the new tooling, demonstrating statistically significant improvements. The benefits were most pronounced in environments where the existing search capabilities were weak, such as with Codex, which routed 91% of search calls through the new tool. The successful configuration is now being prepared for integration into the IntelliJ IDEA MCP Server, with plans for a default rollout in future AI Assistant plugin updates.