CodeLeash: framework for quality agent development, NOT an orchestrator
Hi HN, I built my first project using an LLM in mid-2024. I've been excited ever since. But of course, at some point it all turns into a mess. You see, software is an intricate interwoven collection of tiny details. Good software gets many details right; and does not regress as it gains functionality. My bootstrapped startup, ApprovIQ (https://approviq.com) is trying to break into a mature market with multiple fully featured competitors. I need to get the details right: MVP quality won't sell. So I opted for Test-Driven Development, the classic red/green/refactor. Writing…
In plain words
CodeLeash is a framework designed to help developers build higher-quality AI agents through test-driven development practices. Rather than serving as an orchestrator, it focuses on enforcing rigorous testing and documentation throughout agent development, allowing developers to catch regressions and maintain code quality as functionality grows. The framework is built for teams working on complex AI projects where attention to detail and reliability matter, particularly those in competitive markets where MVP-level quality is insufficient.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
Hi HN, I built my first project using an LLM in mid-2024. I've been excited ever since. But of course, at some point it all turns into a mess. You see, software is an intricate interwoven collection of tiny details. Good software gets many details right; and does not regress as it gains functionality. My bootstrapped startup, ApprovIQ (https://approviq.com) is trying to break into a mature market with multiple fully featured competitors. I need to get the details right: MVP quality won't sell. So I opted for Test-Driven Development, the classic red/green/refactor. Writing tests that fail - then making them pass - forces you to document in your tests every decision that went into the code. This makes it a universal way to construct software. With TDD, you don't need to hold context in your head about how things should work. Your software can work as intricate as you like and still be resilient to regression. Bug in a third-party dependency? Get a failing test, make it pass. Anyone who undoes your fix will see the test fail. At the same time as doing TDD with Claude Code, I also discovered that agents obey all instructions put in front of them! I started to add super-advanced linting: architectural guideline enforcement, scripts that walk the codebase's AST and enforce my architecture, I even added one that enforces only our brand colors in our codebase. That one is great because it prevents agents from picking ugly "AI generic" colors in frontends. Because the check blocks commits with ugly colors, our product looks way less like an AI built it - without human involvement. In time I was no longer in the details of what the agent was building and was mostly supervising the TDD process while it implemented our product. Once that got tedious, I automated that into a state machine too. All the ideas that now allow me build at high quality are in this repo. This isn't your weekend vibe project. I've spent months refining the framework. There are rough edges but it's better out and rough than in hiding until perfect. Hopefully some ideas here help you or your agent. I recommend cloning it and letting your agent have a look! And if you want to contribute please to - and if you want to get in touch, contact details in my profile. Thanks for looking.
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
