AgentFlow Just Got Much Smarter: Introducing Skills & Mandatory Code Review in init-agent
How to get the best out of the agentic tools like Claude Code, Codex, OpenCode etc? Make no mistake, how you set these tools up can GREATLY improve the quality of the code and the success. My AgentFlow - the methodology init-agent sets up, brings software engineering discipline to agentic coding which increases quality of the output.
There is an agent.md for everyone but Claude who names theirs claude.md, but it’s the same role. It’s king of a rule book that applies to ANY task you ask. You can overload this file with information that ends up taking away functionality.
AGENTS.md file served as the "brain," containing every workflow, protocol, and rule the agent needed to know.
But as the methodology evolved, AGENTS.md became a victim of its own success. We kept adding instructions—how to test, how to plan, how to write docs—and eventually hit the context exhaustion threshold. The agents were reading everything, but following nothing.
This paper argues that creating an agent.md increases cost and reduces performance (if written by an LLM), but increases performance if written by a human. https://arxiv.org/abs/2602.11988#:~:text=12%20Feb%202026%5D-,Evaluating%20AGENTS.md:%20Are%20Repository%2DLevel%20Context%20Files%20Helpful,inference%20cost%20by%20over%2020%25.
There is a happy place in the middle and is discussed here: https://www.aibyaakash.com/p/i-open-sourced-claude-code-setup
Today, I’m releasing a massive update to init-agent that solves this fundamentally by introducing Skills Routing, and we're using this new architecture to ship a highly valuable feature: Mandatory Code Review. I am taking the “minimal agent.md that acts as a router to skills”. I’ve update the init-agent to implement this new method and I had some zig bugs fixed: https://github.com/leebase/init-agent
Here is what's new.
1. Skills Routing: The End of Context Bloat
AGENTS.md is no longer an instruction manual; it's a slim router.
Instead of dumping every workflow into one 200-line file, AGENTS.md now acts like an operating system dispatcher.
The actual "how-to" logic has been decoupled into dedicated markdown files inside the skills/ directory.
Why is this so powerful? Cognitive load. When you ask an agent to implement a feature, it no longer loads the "how to groom a backlog" instructions into its working memory. It only loads
development-loop.md. By giving the agent exactly the rules it needs in that moment, adherence to the methodology skyrockets. And for us humans, the system is now infinitely extensible—you can add custom skills to your project without breaking the core OS.
Integrated, Mandatory Code Review - With the new Skills architecture in place, we've finally been able to build a proper Code Review step into the AgentFlow lifecycle.
The sprint Definition of Done in sprint-plan.md now explicitly demands running the Code Review skill.
The output is saved to a newly scaffolded code-reviews/ directory. The sprint-review.md template now points your next fresh AI directly to the review file.
You can no longer close a sprint in an AgentFlow project without an AI doing a full pass and pointing out the edge cases the builder AI missed.
How to Upgrade Your Exisiting Projects: I designed init-agent so that you never have to start over. You can pull the new Skills architecture and updated templates into your existing folders without overwriting your custom logic or data.
Let me know what you think of the flow.
BTW - also interesting is that I monitor x for ideas related to how to use agentic coding tools effectively. I have a standing conversation thread where I refer to them and ask: how are we doing? What good ideas are in this article we should consider implementing.
