Brokk is a new tool for supervising AI coding agents, written in Java.

This talk will cover the principles behind Brokk's design, how Brokk leverages static analysis to help LLMs understand your code at a level deeper than raw text, and Java-specific optimizations like dependency decompilation to avoid hallucinations and JDT pre-build checks.

Brokk is a standalone application because effective context engineering for software development requires a different paradigm than the 40 year old IDE design optimized for reading and writing code at human speeds.

We'll also take a brief look at the Brokk Power Ranking, a benchmark of LLMs' ability to write Java code.

Jfokus

Jonathan Ellis, founder of Brokk (brokk.ai), presents an AI-native code platform designed for large enterprise codebases. Key insights include: model selection strategy (use large models like Opus to decompose tasks, then cheaper mid-size models like Gemini Flash 3 for execution), why existing benchmarks like SWEBench are flawed (over-indexed on Django, susceptible to training contamination), and Brokk's approach of treating code semantically rather than as text. The platform features semantic code search, automatic test execution, dependency ingestion with decompilation support, and AI-assisted code review that leverages session history from the original author. Ellis also recommends leveraging compiler/static analysis tools (like NullAway, ErrorProne) to enforce code quality that prompts alone cannot, and advocates for faster compile times—suggesting Go for greenfield projects.

Brokk: an AI-Native Code Platform, by Jonathan Ellis