Claude Flow, Ruflo and Anthropic Agent Teams: The Claude Multi-Agent Ecosystem

claude-flowrufloagent-teamsanthropicqueen-workermailboxorchestrationcodex-comparison

Claude Flow, Ruflo and Anthropic Agent Teams: The Claude Multi-Agent Ecosystem

The multi-agent story in the Claude ecosystem has evolved rapidly through early 2026. Three distinct layers have emerged: the community-built orchestration platform Ruflo (formerly Claude Flow), Anthropic’s official Agent Teams shipped with Opus 4.6, and a tier of local orchestrators including Claude Squad and OpenClaw+Antfarm. Each occupies a different niche. This article surveys the architecture of each, compares them with Codex CLI’s subagent model, and examines what cross-pollination opportunities exist.

Ruflo: The Community Orchestration Platform

From Claude Flow to Ruflo v3.5

Claude Flow began as a community project by Ruv (ruvnet) to orchestrate multiple Claude Code instances through a queen/worker hierarchy 1. It has since been rebranded as Ruflo and completely rebuilt. The current v3.5 release comprises over 6,066 commits and 250,000 lines of TypeScript and WebAssembly 2.

The core architecture flows through distinct layers:

User → Ruflo (CLI/MCP) → Router → Swarm → Agents → Memory → LLM Providers

Ruflo integrates with Claude Code via MCP, enabling direct command access within Claude Code sessions 2.

Queen-Led Hierarchy and Swarm Topologies

Ruflo employs a queen-led agent hierarchy with strategic, tactical, and adaptive coordinators managing over 100 specialised agent types — coder, tester, reviewer, architect, security, and more 2. Eight worker types (researcher, analyst, documenter, optimiser, etc.) operate beneath these coordinators.

The swarm coordination layer supports four topologies:

graph TD
    subgraph "Ruflo Swarm Topologies"
        direction TB
        M["Mesh<br/>Full peer-to-peer"]
        H["Hierarchical<br/>Queen → Coordinators → Workers"]
        R["Ring<br/>Sequential handoff"]
        S["Star<br/>Central hub routing"]
    end

    subgraph "Consensus Protocols"
        Raft["Raft<br/>Leader-based"]
        BFT["BFT<br/>Byzantine fault tolerant"]
        Gossip["Gossip<br/>Decentralised P2P"]
    end

    M --> Gossip
    H --> Raft
    R --> Gossip
    S --> BFT

Anti-drift protection ensures hierarchical coordinators validate outputs against original goals, preventing agent task deviation 2.

Self-Learning with SONA

Ruflo v3 introduced the RuVector Intelligence Layer, a suite of machine learning components that differentiate it from static orchestration frameworks 3. The Self-Optimising Neural Analysis (SONA) engine learns from every task execution with sub-0.05ms latency, whilst Elastic Weight Consolidation (EWC++) prevents catastrophic forgetting of successful patterns 2.

The platform claims roughly a 250% improvement in effective subscription capacity and a 30–50% reduction in token consumption through intelligent task routing: simple tasks (<1ms) are handled by the Agent Booster (WebAssembly transforms, zero LLM cost), medium-complexity tasks by Haiku/Sonnet, and complex problems by Opus with multi-agent swarms 2. ⚠️ These performance claims come from Ruflo’s own documentation and have not been independently verified.

Adoption

Ruflo reports nearly 100,000 monthly active users across more than 80 countries 3. The npm package remains available under the original claude-flow name 4.

Anthropic Agent Teams: The Official Multi-Agent Layer

Architecture and Enabling

Anthropic shipped Agent Teams as an experimental feature alongside the Opus 4.6 release in February 2026 5. They require Claude Code v2.1.32 or later and are disabled by default — enabled by setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS in settings.json or the environment 6:

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

An agent team consists of four components 6:

Component Role
Team lead Main session that creates the team, spawns teammates, and coordinates work
Teammates Separate Claude Code instances working on assigned tasks
Task list Shared work items that teammates claim and complete
Mailbox Messaging system for direct inter-agent communication

Subagents vs Teammates

The critical distinction from Claude Code subagents is communication topology. Subagents run within a single session and can only report results back to the main agent — they never talk to each other. Agent Teams removes that bottleneck entirely 6:

graph TD
    subgraph "Subagent Model"
        MA[Main Agent] --> SA1[Subagent 1]
        MA --> SA2[Subagent 2]
        MA --> SA3[Subagent 3]
        SA1 -->|results| MA
        SA2 -->|results| MA
        SA3 -->|results| MA
    end

    subgraph "Agent Teams Model"
        TL[Team Lead] --> T1[Teammate 1]
        TL --> T2[Teammate 2]
        TL --> T3[Teammate 3]
        T1 <-->|direct messages| T2
        T2 <-->|direct messages| T3
        T1 <-->|direct messages| T3
        T1 -.->|shared task list| T2
    end

Teammates work independently, each in its own context window, and communicate directly with each other through the mailbox. Task claiming uses file locking to prevent race conditions when multiple teammates try to claim the same task simultaneously 6.

Plan Approval and Quality Gates

Teammates can be required to operate in read-only plan mode until the lead approves their approach. Three hooks enforce quality gates 6:

  • TeammateIdle — runs when a teammate is about to go idle; exit code 2 sends feedback and keeps the teammate working
  • TaskCreated — validates task creation
  • TaskCompleted — validates task completion; exit code 2 prevents premature completion

Display Modes

Agent Teams support two display modes: in-process (all teammates in the main terminal, cycle with Shift+Down) and split panes (each teammate in its own tmux or iTerm2 pane) 6. The default "auto" setting uses split panes if already inside tmux.

Current Limitations

Agent Teams remain experimental with notable constraints: no session resumption with in-process teammates, no nested teams (teammates cannot spawn their own teams), one team per session, and the lead is fixed for the team’s lifetime 6. Token usage scales linearly with teammate count — roughly 4–7× more tokens than single-agent sessions 7.

Tier 2 Local Orchestrators: Claude Squad and OpenClaw+Antfarm

Claude Squad

Claude Squad is a zero-setup tmux-based tool for orchestrating multiple Claude Code agents with a terminal UI dashboard 8. It occupies the simplest end of the local orchestration spectrum — no configuration files, no swarm topologies, just multiple agents visible in terminal panes.

OpenClaw+Antfarm

OpenClaw+Antfarm sits at the more opinionated end. It runs the Ralph loop pattern with planner, developer, verifier, tester, and reviewer roles through one-command installation, using YAML and SQLite for deterministic workflows 8. It represents the “ship while you sleep” paradigm — cron-driven, fully automated.

However, the landscape shifted significantly on 4 April 2026 when Anthropic blocked Claude Pro and Max subscribers from using their subscriptions with third-party agent frameworks including OpenClaw 9. Boris Cherny, Head of Claude Code, stated that subscriptions “weren’t built for the usage patterns of these third-party tools” 9. Users must now use pay-as-you-go billing or direct API keys, with some facing cost increases of up to 50× their previous outlay 10.

This move reinforces the economic advantage of Anthropic’s own Agent Teams over third-party orchestrators for subscription users.

Comparison with Codex CLI Subagents

Codex CLI’s multi-agent model differs architecturally from all three Claude layers:

Feature Codex CLI Subagents Claude Agent Teams Ruflo
Definition TOML-based agent definitions Natural language + subagent defs 100+ pre-built agent types
Addressing Path-based (/root/agent_a) Name-based (mailbox) Swarm topology routing
Communication Report back to parent only Direct peer-to-peer mailbox Queen-coordinated + gossip
Max parallelism max_threads=6, max_depth=1 No hard limit (3–5 recommended) 60+ agents in swarms
Batch spawning spawn_agents_on_csv N/A Swarm deploy command
Sandbox Built-in network/filesystem sandbox Inherits lead permissions Relies on Claude Code sandbox
Protocol Structured messaging (v0.117.0+) Native MCP MCP integration

Where Codex CLI Excels

Codex CLI’s strengths lie in sandbox isolation and token efficiency. Every subagent runs within the same sandboxed environment with network and filesystem restrictions enforced at the process level 11. OpenAI claims Codex CLI uses 4× fewer tokens than comparable approaches 11. The TOML-based agent definition system offers deterministic, version-controllable agent configuration:

[agent.security-reviewer]
model = "o3"
instructions = "Review code for security vulnerabilities"
tools = ["read", "grep", "glob"]
max_threads = 2

Where Claude Agent Teams Excel

Agent Teams’ mailbox system enables a fundamentally different coordination pattern. In a real-world Next.js migration example, an API refactoring agent discovered a type change that would break the frontend and flagged it directly to the frontend agent, which adjusted its approach without manual orchestration 7. This kind of cross-agent coordination simply doesn’t exist in Codex CLI’s report-back-to-parent model.

The plan approval mechanism also has no direct equivalent in Codex CLI — teammates can be held in read-only mode until the lead explicitly approves their approach, providing a safety gate for high-risk changes.

Cross-Pollination Opportunities

Several patterns could transfer between ecosystems:

  1. Mailbox for Codex — Codex CLI’s subagents could benefit from peer-to-peer messaging rather than routing everything through the parent. This would enable the kind of cross-agent discovery that Agent Teams demonstrate.
  2. TOML definitions for Claude — Agent Teams currently rely on natural language prompts or subagent definitions to configure teammates. Codex’s declarative TOML approach offers more reproducible team configurations.
  3. CSV batch spawning for Claude — Codex CLI’s spawn_agents_on_csv pattern, which creates agents from spreadsheet data, has no equivalent in Agent Teams and could enable data-driven team composition.
  4. Ruflo’s self-learning for both — Neither Codex CLI nor Agent Teams learn from previous task executions. Ruflo’s SONA pattern of reinforcing successful agent routing could improve both.

The Emerging Architecture

The three-layer Claude multi-agent ecosystem maps to distinct use cases:

graph LR
    subgraph "Lightweight"
        CS[Claude Squad<br/>Zero-setup tmux]
    end

    subgraph "Session-Scoped"
        SA[Subagents<br/>Single session, report back]
        AT[Agent Teams<br/>Multi-session, peer-to-peer]
    end

    subgraph "Platform-Scale"
        RF[Ruflo<br/>100+ agents, swarm topologies]
        OC[OpenClaw+Antfarm<br/>Deterministic YAML workflows]
    end

    CS -->|"need coordination"| AT
    SA -->|"need peer messaging"| AT
    AT -->|"need 60+ agents"| RF
    AT -->|"need cron automation"| OC

For Codex CLI users, the key takeaway is architectural: the future of multi-agent coding tools is moving from hub-and-spoke (parent coordinates everything) toward peer-to-peer (agents coordinate directly). Whether that arrives in Codex CLI via A2A protocol support, an MCP-based mailbox, or a native implementation remains to be seen — but the direction is clear.

Citations

  1. Ruflo GitHub repository — ruvnet/ruflo ↩

  2. Ruflo README and technical documentation ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. Analytics Vidhya — Claude Flow: The AI Orchestration Framework Redefining Multi-Agent Automation ↩ ↩2

  4. claude-flow npm package ↩

  5. TechCrunch — Anthropic releases Opus 4.6 with new ‘agent teams’ ↩

  6. Claude Code Docs — Orchestrate teams of Claude Code sessions ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  7. Particula — Codex vs Claude Code: Which CLI Agent Wins for Your Workflow in 2026 ↩ ↩2

  8. Addy Osmani — Orchestrating Coding Agents, O’Reilly CodeCon 2026 ↩ ↩2

  9. VentureBeat — Anthropic cuts off the ability to use Claude subscriptions with OpenClaw ↩ ↩2

  10. TechCrunch — Anthropic says Claude Code subscribers will need to pay extra for OpenClaw usage ↩

  11. OpenAI Codex CLI documentation and release notes ↩ ↩2