Ruflo

Ruflo has the clearest definition of this whole category anyone has written: "Agent = Model + Harness. The model writes; the harness gives it tools, memory, loops, sandboxes, and controls so it can actually work." That sentence is worth the visit on its own. It also calls itself "the original agent harness", which is the one claim in the README that does not survive a date check. At 73,553 stars and MIT, it is the multi-agent end of this series: swarms, consensus and a memory layer, rather than one agent editing one repository.

About that claim

Ruflo's repository was created on 2 June 2025, under its former name Claude Flow, and renamed by its author rUv. That is genuinely early for this space, well ahead of DeepSeek Harness in August 2026 or ZCode in September.

It is not first. OpenCode was created on 30 April 2025, a month earlier, and the frameworks the README itself names as reference points, LangGraph, AutoGen and CrewAI, predate both. The README offers no timeline or precedence argument to support the word "original", and we could not construct one.

Call it early rather than original. The distinction matters because the rest of the project is strong enough not to need the superlative.

Swarms, concretely

Most multi-agent frameworks mean "we call several models in a loop". Ruflo means something more structured: hierarchical, mesh and adaptive topologies with consensus, organised as a Queen-led hierarchy running Raft, Byzantine or Gossip protocols, with behavioural trust scoring and task routing on top.

Borrowing Byzantine fault tolerance for agent coordination is a real design decision rather than a metaphor. It assumes some of your agents will produce bad output and builds agreement that tolerates it, which is a more honest model of how LLM agents behave than assuming every result is usable.

Memory is an HNSW-indexed vector database the project calls AgentDB, with sub-millisecond retrieval and a reported 1.9x to 4.7x speedup over brute-force search at certain dataset sizes. Learning is handled by what it calls SONA neural patterns, ReasoningBank and trajectory learning, aimed at retaining successful patterns across sessions.

Those learning components are the least documented part and we would treat them as the least verified. The claim is the same shape as Hermes's learning loop, and neither has independent evaluation.

Two install paths, and they are very different

curl -fsSL https://cdn.jsdelivr.net/gh/ruvnet/ruflo@main/scripts/install.sh | bash

npx ruflo@latest init wizard
npx ruflo@latest init

claude mcp add claude-flow -- npx ruflo@latest mcp start

The scope difference is worth understanding before you pick. The Claude Code plugin path gives you slash commands and agent definitions with zero workspace files. The CLI path installs the full harness: 35 plugins and roughly 98 agents, against a shipped library of more than 100 specialised agents for coding, testing, security, docs and architecture.

That is a large footprint, and running npx ruflo init in a project expecting a light tool will surprise you. Start with the plugin path.

Note also that the MCP registration still uses the name claude-flow, a leftover from the rename. Harmless, but if you are searching your own MCP config later, that is the string to look for.

Where it sits

The series now covers six layers, and ruflo occupies one nothing else here does.

Ruflo also brings its own memory and learning, which puts it in the same overlapping-responsibilities position as Hermes. Run it alongside claude-mem and you have two systems with opinions about what to remember. Nobody has written the guidance for combining these yet, which is itself a sign of how new the stack is.

The community test

We have been checking whether a project accepts bug reports as a proxy for who it is built for. Ruflo has Issues enabled with 1,030 open against 8,728 forks, which puts it with OpenCode, Herdr and Hermes on the open side, and apart from every harness built by a lab that sells model access.

1,030 open issues on a project this size is a lot, and as with Hermes we cannot tell engagement from backlog at a distance. But the tracker is open, which is the part you can verify.

Who should use it

Use ruflo if you actually need multiple agents coordinating rather than one agent working faster. Parallel review across a large codebase, pipelines where specialised agents hand off, anything where consensus over several attempts beats trusting one. The topology and trust-scoring machinery is real engineering and there is little else like it at this license.

Do not use it as a better single coding agent. That is what the harnesses in this series are for, and ruflo's footprint is hard to justify if you are running one agent in one repository. And treat the self-learning claims as unproven until somebody outside the project measures them.

Sources and further reading

• Ruflo on GitHub
• Cognitum.One, the project site
• OpenCode, created a month earlier

Ten minutes: install the Claude Code plugin path rather than the full CLI, and give a swarm a task you would normally split across two sessions yourself, such as reviewing a change for both security and performance. If the consensus output beats what one agent gave you, the coordination machinery is earning its footprint. If it just costs four times as many tokens to reach the same answer, it is not.

Tested on: not independently tested. Architecture, topology, memory and learning descriptions are quoted from the project's own README, and the performance figure for AgentDB is author-reported. The precedence check on "the original agent harness" uses repository creation dates from the GitHub API. Star, fork and issue counts were read from the same API on the date below.
Date checked: 2026-09-30