Netflix Cuts AI Coding Agent Costs With JetBrains IDE Intelligence

The problem: Java and Kotlin code behind JAR files

Netflix's Application Development Platform (ADP), GenAI Platform and Agentic SDLC teams found that coding agents struggled to navigate Java and Kotlin code, especially across compiled library (JAR) boundaries. Paul Bakker, Staff Software Engineer at Netflix, said in the JetBrains write-up that "Coding agents like Claude Code are slow and unreliable at navigating Java/Kotlin code, especially across library (JAR) boundaries."

The cause, as JetBrains describes it, is that JVM projects typically ship libraries as JAR files only, never as source. Agents have to unzip the JARs, guessing which ones matter, potentially decompile class files, and then grep through an incomplete view of the code.

How the setup works

JetBrains says its code intelligence engine is based on a deterministic index of a project's source files, including all libraries. The intelligence is now exposed through an LSP (language server) and is available in VS Code-compatible editors. Netflix worked with a headless build of that engine, so no IDE had to be open.

An LSP alone is not enough for an agent, according to JetBrains. It is a server speaking a protocol, with no easy way for an agent to call it and no built-in process management to reuse a running instance. Agents also call it in many small steps, and starting a fresh server each time would add a multi-second cold-start cost.

JetBrains says it is working on a daemon, but none was available at the time of evaluation, so Netflix built its own daemon and CLI. The daemon's session manager routes each call by project root to a matching server process that holds the project's index in memory. A second call on the same project skips indexing and goes to a warm server over stdio. Both tools are written in Java and distributed as a GraalVM executable, which avoids JVM startup and runtime overhead.

How Netflix measured the gains

The benchmark kept the model, task and codebase the same. The only variable was whether the agent could use JetBrains IDE intelligence. Netflix used two real codebases: an internal medium-size service using Spring Boot 4 and Gradle, and its open-source DGS Framework, a mixed Kotlin and Java GraphQL framework. Runs took place in July and August 2026.

Tasks fell into two groups: small, focused navigation questions such as where an interface is used or which implementations exist, and full end-to-end work such as implementing a feature, refactoring a class or removing code. The models tested were Claude Code with Opus 5 and Sonnet 5, and Codex with GPT-5.6 Sol, each at medium and high effort.

Token use and time to completion were the main metrics. An LLM acted as a judge for correctness. Tasks that crashed, hit a budget cap or thrashed in either arm were excluded from both.

Results and caveats

According to JetBrains, every model, agent and task type became more efficient, typically by 20, 30%, and by up to about 65% in the best case. The strongest model at high effort gained most: about 65% on code navigation and about 24% on end-to-end tasks. A mid model at high effort saw about 29% and about 30%. Correctness held steady or improved in every run.

• On code navigation, every configuration used fewer tokens, with savings from roughly 15% to 60%.
• One tool and model combination batches multiple search calls into a single command, which hid any time benefit on navigation tasks. JetBrains says the end-to-end suite is the more reliable read on time, where the tool was faster across every model and agent tested.

JetBrains stresses that the figures are a snapshot and will shift with new model releases, which is why it reports ranges instead of per-model results.

What it means

For teams running coding agents on JVM projects, the results suggest that giving agents a proper code index can reduce token spend and waiting time. JetBrains notes that the stronger models are more expensive, so efficiency gains matter more there, and that these models are also better at using indexes effectively. The daemon and CLI were built by Netflix for its own needs, while JetBrains says it is working on delivering a daemon of its own.