'Han' workflow runs multiple Claude agents in parallel using a spec-first approach

Han is a community-shared workflow that coordinates several Claude Code agents like a small team, each following a written specification. Instead of asking one AI to do everything, you write a plan first, then agents use it as their source of truth. It's a practical method for solo developers who want more reliable results from Claude Code.

Most people use Claude Code by chatting with a single AI and iterating. Han introduces an 'agent swarm' model: you split the work among several Claude agents that each handle a specific role — writing code, reviewing it, checking against requirements, and so on. The key starting point is writing a clear specification (a document describing exactly what you want to build) before any code is written.

The 'evidence-based' part means agents don't improvise — they refer back to the spec to make decisions and verify their output. This reduces drift and keeps a complex feature on track even when multiple agents are running. This is a user-shared workflow posted on Reddit, not an official Anthropic product, so it requires hands-on setup and adaptation to your own project.

Key points

  • Write a detailed spec (requirements doc) before coding — agents use it as ground truth
  • Multiple Claude agents divide roles: one writes code, another reviews, etc.
  • 'Evidence-based' means agents check their work against the spec, not just intuition
  • Designed for solo developers using Claude Code on non-trivial features
  • Community workflow, not an official tool — expect manual setup and customization

Quick term guide

workflow
A repeatable set of steps for getting a task done.
agents
AI helpers that follow your instructions and make changes for you.
specification
A written document that defines exact rules a piece of software must follow
developers
Developers are people who build software, apps, or websites.
iterating
Going through repeated small rounds of editing and testing to improve something step by step.
agent swarm
Multiple AI agents working at the same time, each assigned a different task.
swarm
A group of AI agents each handling part of a task at the same time, like a team instead of one person.
evidence-based
Making decisions by referring to a fixed document or data rather than guessing.
Read original