Three frameworks for autonomous Claude Code. Tested each one. They all work. The problem is something else.
The last two weeks I dug into wrappers for Claude Code. Simple goal on paper - make the agent run on its own: plan, decompose, execute, document. In practice, harder than it sounds.
What I tested
Claude Flow (by Reuven Cohen / ruvnet). 13.8k stars on GitHub. Positions itself as an enterprise platform for multi-agent orchestration. 60+ specialized agents out of the box. Swarm coordination. Memory. RAG integration. V3 added a self-learning architecture called SONA.
Impressive on paper. In practice - the first 2 hours went to figuring out which agents I actually needed and which were overkill. Claude Flow is a Swiss Army knife. Powerful. But sometimes you just need a screwdriver.
spec-kit (from GitHub). 68.4k stars. Spec-driven development - specs become executable. Five main commands: /speckit.constitution (principles), /speckit.specify (requirements), /speckit.plan (plan), /speckit.tasks (tasks), /speckit.implement (implementation).
Strict approach. First the project constitution. Then the spec. Then the plan. Then tasks. Then code. Sounds like waterfall? A bit. But for large features - exactly right. For a quick prototype - too much ceremony.
Superpowers (by Jesse Vincent / obra). 47.7k stars. v4.1.1. TDD-first philosophy. Brainstorm - plan - execute. If Claude writes code before tests - Superpowers deletes it and makes it start over.
Harsh? Yeah. But 85-95% test coverage is no joke. Install via the Claude Code plugin marketplace. Of the three frameworks - the most opinionated. Either accept its rules or don’t use it.
What I got
Spent a week with each. Wrote posts, automations, small projects.
Claude Flow - powerful when a task needs multiple agents at once. Research + writing + review in parallel. But setup eats time.
spec-kit - best for large features where decomposition matters. Plan - tasks - implement runs predictably. GitHub knows how to build DevTools.
Superpowers - the strictest one. Forces you to think before coding. TDD out of the box. For production code - ideal. For creative tasks - constraining.
What I realized
Originally wanted to write a post: “most frameworks don’t work.” After a week of testing - changed my mind.
All three work. Seriously. Each has its own approach. Claude Flow - for orchestration. spec-kit - for specifications. Superpowers - for discipline.
The real question isn’t “which framework to pick.” The real question is whether you understand how to manage autonomous agents. What to put in context. When to step in. When to let go.
I’m still not sure I’ve found the ideal approach. I mix them. spec-kit for planning. Superpowers for execution. Claude Flow - when I need heavy artillery. And frameworks pair well with custom Claude Code commands - your own slash-commands for recurring tasks.
A framework is a steering wheel. But you’re still pressing the pedals.