What Claude Flow Was

Right, so this tool landed in mid-2025 and it properly turned heads. Claude Flow was built by Reuven Cohen (GitHub handle ruvnet), and the pitch was straightforward: take Anthropic's Claude, strap a multi-agent orchestration layer on top of it, and let coordinated swarms of AI agents tackle software development tasks. Not one agent fumbling through your codebase on its own. Dozens of them, working in parallel, with a queen agent calling the shots.

The repo went live on GitHub around June 2025 under ruvnet/claude-flow, and the thing that set it apart from other "run multiple agents" tools at the time was SPARC.

SPARC: The Methodology Baked Into the Tool

SPARC stood for Specification, Pseudocode, Architecture, Refinement, Completion. It was a structured, test-driven approach to AI development that came packaged directly into Claude Flow. You didn't have to invent a workflow or figure out how to prompt your agents. You ran npx claude-flow sparc run and the tool walked your swarm through each phase.

Ten specialised modes. The agents knew what phase they were in, what their job was at each stage, and how to hand off work to the next phase. If you'd been struggling with the "I told the agent to build a feature and it went off the rails" problem, SPARC was the answer Claude Flow offered. Structure the work, structure the agents, structure the output.

How the Swarm Actually Worked

This is where it got interesting. Claude Flow used what they called a "hive mind" architecture. A queen agent sat at the top, coordinating sub-agents below. Each sub-agent had a specialisation. Some did research. Some wrote code. Some reviewed it. The queen figured out who should do what and when.

By the time v2.7 rolled around, the platform had:

• 60+ specialised agents running in coordinated swarms
• AgentDB memory powered by SQLite with semantic queries
• Neural memory enhancement for cross-session recall
• Claude Code integration via MCP (Model Context Protocol)
• Parallel agent execution across tasks

The AgentDB bit was particularly clever. v2.7.x introduced 150x faster semantic queries with 56% less memory usage. So your agents could actually remember what they'd done across sessions without the whole thing grinding to a halt.

What Happened Next

Claude Flow eventually became Ruflo in early 2026. Not just a rename, a full architectural shift to Rust/WASM. But that's a separate story.

📚 Geek Corner

SPARC to Skills: The original SPARC methodology was rigid by design. Five phases, sequential, with explicit gates between them. This works brilliantly for well-defined tasks but fights against the reality that most development is messy and iterative. Ruflo's shift to skills-based orchestration reflects a pattern seen across the whole AI tooling space in 2025/2026: structured phases give way to composable capabilities. HumanLayer's RPI methodology made the same move. GitHub's spec-driven development is another flavour of it. The industry is converging on "give agents the right tools and context, then let them figure out the order" over "prescribe the exact sequence." SPARC was the right answer for mid-2025. Skills are the right answer for where we are now.

When to Use It

If you're running Claude Code and you want to throw multiple agents at a problem in parallel with proper coordination, Ruflo (the current incarnation) is one of the most mature options out there. 29,000+ stars on GitHub. Active development. Real architectural investment under the hood.

The tradeoff is complexity. This is not a "pip install and go" tool. It's an enterprise-grade platform with consensus algorithms (Raft, Byzantine, Gossip), distributed swarm intelligence, and a WASM runtime. If you need five agents working on different files simultaneously with shared memory and coordination, it's brilliant. If you need one agent to fix a bug, it's massive overkill.

For simpler multi-agent needs, have a look at Claude Squad which takes a much lighter approach to running parallel agents.