AgenticDev: a multi-LLM framework for generating tested code

In late 2025, after spending hours prompting LLMs one by one to generate code, a question kept nagging me: what if multiple LLM agents could collaborate to produce a complete project? Not a single agent doing everything, but a specialized team — an architect, a developer, a tester — each with its own role, tools, and constraints. That's how AgenticDev was born, a Python framework that orchestrates 4 LLM agents to turn a plain-text request into tested, documented code. In this article, I share the architecture decisions, the problems I ran into, and the lessons learned. Starting point: testing the limits of multi-agent collaboration My initial goal was simple: explore how far LLM agents can collaborate autonomously. Not a throwaway POC, but a real pipeline where each agent has a clear responsibility: Architect — analyzes the request and produces a technical specification ( spec.md ) Designer — generates SVG assets from the spec Developer — implements the code following the spec and inte