Brokk: an AI-Native Code Platform
summary
Brokk is a Java standalone app that supervises AI coding agents using static analysis and dependency decompilation so models see parsed APIs and decompiled libs rather than raw text. Benchmarks (Brokk Power Ranking) show midsize models like Flash 3 are fastest and cheapest, Opus wins accuracy but costs ~$5 per million tokens vs Flash at ~$0.50, and midsize success rates collapse as task length doubles (small models fall to single digits at 100k context, large models degrade to ~80–85%). Built features include semantic code search, enforced test runs, errorprone integration (a redundant-null patch was upstreamed), and a research→plan→build loots mode for task decomposition.
AI-written summary · about
chapters (11)
- 0:00Introduction to AI for code
- 1:02Choosing AI model for coding
- 8:56Flaws in coding benchmarks
- 10:43Verification tools with AI
- 14:23Investing in developer tools
- 15:30Enterprise Java and AI tools
- 18:16Live demo of Brock’s visualization
- 29:27Handling dependencies and decompiled code
- 30:10AI-assisted code review
- 33:01Human review in AI coding
- 37:37Q&A on Brock’s data and models
related talks
53mBrokk, an AI Native Code Platform for Java, in Java
35mUsing coding agents with open models
24mBeyond Code Completion: The Shift of AI Programming from Pair to Peer
18mFast Models Need Slow Developers
15mUsing AI effectively in large Java codebases
22mSponsored Demo: Cut the Bill, Not the Brains. AI Coding Agents That Ship More and...
48mHow to Verify AI-Generated Code with Codex
13mSWE-Marathon: Evaluating Coding Agents at Billion-Token Scale