agent-flow

Claude Agent Flow

A Claude Code plugin that brings a multi-agent development pipeline to your repository — adversarial review loops, 3 agents, 28 skills, and 14 slash commands.

Full documentation → · Jump to installation ↓

Note: Install this into your existing project using the install.sh script: Jump to installation ↓ For a fresh start or sandbox/workspace use, you can also clone this repo directly and open it in Claude Code — run /install once to initialise sync state.

Why use this framework/Claude Code plugin and why did I build it

I started this project because if I am honest I was a bit disappointed with the current crop of AI tooling. As amazing as their results often are, when the model says "I am done", it was always far from done. Maybe you asked for a mobile friendly web page and it turned out that the result wasnt in fact mobile friendly. Or maybe you asked it to fix some code and then on running the tests, you saw that it had made introduced another bug. Either way I was spending too much time hand holding and guiding the models.

I am a big Claude Code user and it supports so many great things so I thought why don't I try and build a workflow that would help me formulate my idea, research it for me, sanity check it and then build it for me all without my intervention. Obviously I would need to discuss with the AI what I wanted but once we had a good plan, I wanted my work to be turned into a "ticket" and then have the AI build it for me and not come back to me until it was "Done Done"!

But the big part was once it thought it was done, it would need to get the approval of multiple separate AI agents that would be very thorough and check it's work. "Works on mobile?", then prove it with screenshots! "Code looks good", then lets run all the tests. If anything breaks, send it back to the agent that did the work for correction. Only when the entire flow if complete and every agent is satisfied, is the job "Ready for Review"

Today it's 3 core agents, 28 skills, and a full CI/CD integration — built entirely by running itself. What began as a simple automation became a full on factory line capable of solving multiple tickets at once. When you couple it with Claude Code for Web you can literally have AI build your ideas whilst you are out and about. It gets rather addictive :-)

The Philosophy

Two ideas drive everything in this project:

Quality emerges from adversarial loops. A dedicated adversary that actively tries to break the code finds what a reviewer never sees. The critic's job is not to approve — it is to fail the work. The loop runs unattended until nothing breaks, compressing days of real-team code review into minutes. This is how you catch race conditions, silent failures, and injection vectors that slip past conventional review.

Specialisation beats generalism. A single agent that must design, implement, test, and review the same work carries compounding bias. Separation of roles removes the tension between building and verifying. Specialised agents and skill-backed roles each own exactly one part of the problem — clean context, no role confusion. The Critic's sole mandate is to break the work; the Explorer maps without building; the Researcher validates without implementing.

The full reasoning behind these choices is on the Why Agent Flow page.

How It Works

You describe what you want. The pipeline handles the rest.

/plan — Socratic conversation that asks questions, explores approaches, produces a structured brief with acceptance criteria saved as a plan file in your repo.

/build — Takes a plan file and runs the full pipeline: Explorer maps codebase, the architect skill designs the approach, builders implement, then the adversarial Critic loop kicks in (Critic tries to break it, builder fixes, repeat until PASS), then the test suite runs, the review skill does a final pass, and the author skill updates docs. Runs unattended.

/review — For code you wrote yourself without using /build. Runs the adversarial critic loop to stress-test your diff before merging.