Skip to content

Netflix cut AI coding agent tokens 20-30% with JetBrains index

JetBrains reports that plugging its headless IntelliJ code index into Claude Code and Codex made agents cheaper and faster on Java and Kotlin work, with the biggest gains on the strongest models.

Draftship Editorial4 min read

Netflix wired JetBrains code intelligence into its coding agents

JetBrains published a customer story on 15 October 2026 describing how Netflix added JetBrains IDE intelligence for coding agents to Claude Code and Codex, then benchmarked the result on real codebases. Netflix's Application Development Platform, GenAI Platform and Agentic SDLC teams ran the evaluation in July and August 2026. JetBrains reports agents were typically 20% to 30% more efficient on tokens and time, and up to about 65% in the best case. The numbers are in JetBrains' write-up of the Netflix evaluation.

The problem: libraries without source

JVM projects usually ship libraries as JAR files with no source attached. As the post describes it, an agent in such a project has to unzip the JARs it guesses matter, possibly decompile class files, then grep through that incomplete view. Paul Bakker, a Staff Software Engineer at Netflix, is quoted calling agents like Claude Code slow and unreliable at navigating Java and Kotlin code, especially across library boundaries. The fix is the code-indexing engine behind IntelliJ IDEA: a deterministic index of project source files that includes all libraries, exposed over an LSP and usable in VS Code-compatible editors. Netflix ran a headless build with no IDE open.

The plumbing Netflix had to build

An LSP server alone is awkward for an agent. It speaks a protocol, offers no convenient way for an agent to call it and manages no processes. Indexing also takes real time, and a realistic task makes many small calls, so cold-starting a server each time would add multi-second waits even on a project indexed moments earlier.

Netflix wrote a daemon and a CLI. Tool calls are short-lived and stateless, shelling out to the CLI, while the daemon stays up and routes each call by project root to the server process holding that project's index in memory. A second call on the same project skips indexing and reaches a warm server over stdio. Both are Java, shipped as a GraalVM executable so the JVM adds no startup cost.

What the benchmark measured

Two codebases: an internal, medium-sized Spring Boot 4 and Gradle service, and Netflix's open-source DGS Framework, a Kotlin and Java GraphQL framework. Tasks were either focused navigation questions, such as where an interface is used or which implementations exist, or end-to-end work such as adding a feature, refactoring or removing code. Configurations covered Claude Code with Opus 5, Claude Code with Sonnet 5 and Codex with GPT-5.6 Sol at medium and high effort. Tokens and time to completion were the main metrics, with correctness checked against a reference change set and by deterministic scoring on navigation.

Average gain was about 25%. The strongest model at high effort showed roughly 65% on navigation and 24% end to end. A mid model at high effort: about 29% and 30%. At medium effort: about 24% and 22%. A lighter configuration: about 17% on tokens only, since batched search calls made the time comparison unrepresentative, and 21% end to end. Navigation token savings ranged from about 15% to 60%, every end-to-end combination was roughly 10% to 30% more efficient, and correctness held steady or improved.

Limits and open questions

The figures are a snapshot, JetBrains says, because each new model release changes how agents search and use tools, which is why it gives ranges rather than per-model results. Only Java and Kotlin codebases were tested, so nothing here speaks to other languages. Netflix rejected two alternatives first: an internal CLI talking to a running IntelliJ IDEA instance, which needed an IDE and did not suit headless agents in CI and similar environments, and an existing Eclipse LSP, which did not work with its Gradle-based build. The post does not say whether JetBrains will ship its own daemon, which was not available at evaluation time, or whether the benchmark harness is public. Netflix's conclusion, as quoted: "the benchmarks show that there is a benefit to having a deterministic intelligence layer alongside your agent".

What it means for people directing agents

Handing an agent precise structural knowledge of a codebase cuts cost and time, and the effect grows with stronger models run at higher effort. Draftship is an AI app builder where you describe an app in plain language and it writes, builds and hosts the code, with a built-in database and user sign-in. Paid plans add private projects and a choice of AI model, covered in using the AI chat and choosing a model. Draftship is not affiliated with JetBrains or Netflix.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.