Coding agents stop unzipping JARs: Netflix wires JetBrains' code engine into Claude Code and Codex, cutting tokens and time by 20–30%

Coding agents gain a direct channel to query code structure in Java/Kotlin projects, instead of reconstructing library understanding by unzipping JARs, decompiling class files and grepping. Libraries in JVM projects typically ship as JAR files without source, so agents had to guess which ones mattered and work from an incomplete decompiled view, making navigation across library boundaries slow and unreliable. Each comparison held the model, task and codebase fixed and varied only whether the agent could invoke JetBrains code intelligence; every model, agent and task type tested came out more efficient.

The approach runs IntelliJ IDEA's deterministic code index headlessly and adds a daemon and CLI that let an agent call it. JetBrains already exposes that index through an LSP, but an LSP is only a server speaking a protocol, with no built-in process management for reusing a running instance between calls. Netflix's daemon routes each call by project root to the server process holding that project's in-memory index, so a second call skips indexing and goes straight to a warm server; daemon and CLI are written in Java and shipped as a GraalVM executable to avoid JVM startup overhead.

The results come from two real codebases: a medium-size internal service using Spring Boot 4 and Gradle, and Netflix's open-source DGS Framework, a mixed Kotlin/Java GraphQL framework. Tasks cover small code-navigation questions and end-to-end work such as implementing a feature, refactoring a class or removing code. Evaluation ran July–August 2026 with Claude Code on Opus 5 and Sonnet 5 and Codex on GPT-5.6 Sol, each at medium and high effort. Netflix plans to run the same benchmark methodology against Python, TypeScript/JavaScript and Go, and JetBrains says it will continue the collaboration and support the evaluation.

The write-up calls the percentages a snapshot rather than a fixed benchmark result, since each new model release changes how agents search, reason and use tools, which is why it reports ranges instead of per-model numbers. In code navigation, one tool and model combination already batches multiple search calls into one, so that combination showed no time benefit, which the report attributes to the setup rather than real-world use. JetBrains notes that the LSP's basic primitive is a text position, a fragile identity for agents, and the final agent-facing abstraction is still undecided.