Netflix Boosts AI Coding Agent Efficiency with Headless JetBrains Intelligence
Integrating headless IDE engines improves agent performance by up to 65% in Java projects.
1The Challenge of JAR Boundaries
Coding agents often struggle with Java and Kotlin projects because dependencies are distributed as JAR files rather than source code. Agents must manually decompile or grep through these files, leading to unreliable navigation. Netflix identified this as a major bottleneck for their Agentic SDLC teams, hindering the effectiveness of automated coding tasks.
2Headless IDE Integration Strategy
Netflix integrated the headless code-intelligence engine from JetBrains, the same technology powering IntelliJ IDEA, into their agents. By running this engine without a graphical interface, they provided agents with deterministic, index-based code understanding. This setup allows agents to query complex project structures directly, bypassing the limitations of standard text-based search methods.
3Infrastructure for Agentic Workflows
Because the Language Server Protocol (LSP) lacks built-in process management, Netflix developed a custom daemon and CLI. This infrastructure keeps the project index warm in memory, preventing multi-second cold-start penalties during repeated agent calls. The solution is distributed as a GraalVM executable, ensuring minimal runtime overhead for the JVM-based development environment.
4Benchmarking Methodology and Scope
Netflix evaluated performance across two codebases: an internal Spring Boot 4 service and their open-source DGS Framework. They tested Claude Code and Codex models using medium-to-high effort settings. The benchmark measured token consumption and time-to-completion, while an LLM-as-a-judge verified the correctness of the generated code changes against reference sets.
5Efficiency Gains Across Models
Results showed a consistent efficiency improvement of 20–30% on average, with peak gains reaching 65% in specific tasks. Stronger models demonstrated the most significant improvements, as they are better equipped to leverage the deterministic index provided by the JetBrains engine. These gains were observed without any degradation in the final code correctness.
6Navigation vs. End-to-End Tasks
For focused navigation tasks, such as finding interface implementations, agents used significantly fewer tokens. In end-to-end tasks like refactoring or feature implementation, the efficiency boost stemmed from requiring fewer tool calls to reach accurate conclusions. The broader end-to-end suite proved to be the most reliable metric for measuring real-world time savings.
7Rejected Alternative Approaches
Netflix initially considered connecting agents to a running IntelliJ IDEA instance, but this required an active GUI, making it unsuitable for headless CI or distributed environments. They also evaluated the existing Eclipse LSP used in VS Code, but it failed to support the specific requirements of Netflix’s Gradle-based build system architecture.
8The Role of Stronger Models
While all tested models became more efficient, the largest gains were concentrated in high-effort, stronger model configurations. These models are typically more expensive to run, making the efficiency improvements financially and operationally significant. The data suggests that providing high-quality context allows these advanced models to reason more effectively over large codebases.
9Correctness and Stability Metrics
Throughout the evaluation, correctness remained steady or improved across every model and agent combination. Netflix implemented strict outlier handling, excluding tasks that crashed or hit budget caps to ensure the comparison remained fair. This rigorous approach confirmed that the efficiency gains were not achieved at the expense of code quality or reliability.
10Future Language Expansion Plans
Following the success with Java and Kotlin, Netflix plans to apply this benchmark methodology to Python, TypeScript, and Go. Each language presents unique challenges regarding library management and dependency resolution. The team aims to adapt the headless engine approach to these ecosystems, recognizing that deterministic intelligence is essential for scaling agentic workflows.
11JetBrains' Evolving LSP Vision
JetBrains has shipped this intelligence as part of the Java Kotlin by IntelliJ IDEA extension. They acknowledge that the standard LSP, designed for human cursor-based interaction, is insufficient for AI agents. Future developments will focus on evolving the protocol to better support agentic queries, moving beyond simple text-document position primitives.
12Deterministic Intelligence as Foundation
The experiment proves that pairing LLMs with a deterministic intelligence layer significantly enhances agent performance. By solving the JAR boundary problem, Netflix has established a scalable pattern for AI-driven development. This integration serves as a blueprint for organizations looking to increase the ROI of their coding agents in complex, enterprise-grade environments.