SWORDSwarm — Summary

SWORDSwarm (v42.0) is a Python-based multi-agent orchestration system with 117 specialized agent definition files (in YAML+Markdown format) organized in a 4-level corporate hierarchy (Executive/Senior Management/Middle Management/Specialists), hardware acceleration via Intel NPU and AVX2/AVX-512 SIMD, a three-tier architecture (Binary/C+Rust → Hook/Python → Agent/Markdown layer), OpenAI Codex integration, and a "SWORD" invocation model where Claude Code agents can invoke orchestrators via a Python subprocess bridge.

Problem it solved: Complex enterprise AI orchestration requiring many specialized agents with clear authority chains, hardware-accelerated execution, and the ability to bolt on additional tooling (Warp terminal, Cursor, Windsurf, VS Code) via config files.

Distinctive traits: (1) 117 specialized agent .md files in a corporate hierarchy (DIRECTOR, CSO, LEADENGINEER, AGENTSMITH, ARCHITECT, etc.); (2) Intel NPU + AVX2/AVX-512 hardware acceleration (7-10x speedup claimed); (3) Three-tier architecture: C/Rust binary layer → Python hook layer → agent markdown layer; (4) Multi-IDE support via config files (.cursor, .windsurf, .warp, .vscode); (5) Described with extreme self-promotional language ("production-ready", "enterprise-grade", "military-grade optimization").

Credibility note: Many claims (82% test coverage, 95% on-time delivery rate, 98.7% deployment success rate) appear aspirational rather than verified. The repo has 24 stars and was last committed in February 2026.

differs_from_seeds: SWORDSwarm's agent definition format (YAML+Markdown with UUID, metadata, version, proactive_triggers) is closest to BMAD-METHOD's persona-md pattern but much more verbose and structured. Unlike BMAD (34 skills), SWORDSwarm has 117 agent files with explicit corporate hierarchy. No other framework in this corpus attempts hardware acceleration for agent execution.

SWORDSwarm — Overview

Origin

Created by SWORDOps (GitHub org: SWORDOps, alt: SWORDIntel). No license file found in GitHub API response. Default branch: maine (unusual). Version: v42.0 (from README). 24 stars, last committed February 2026.

Philosophy

From README:

"Yeet 88 agents at a problem and see what survives."

And more formally:

"Production-ready multi-agent AI orchestration system with hardware acceleration and AI-powered development tools." "Claude Agent Framework is an enterprise-grade platform for building intelligent, coordinated agent systems with unprecedented performance through Intel NPU acceleration."

The README oscillates between satirical humor ("yeet 88 agents", "i bolted warp on the side poorly") and extreme enterprise marketing claims ("military-grade optimization", "zero vulnerabilities").

Corporate hierarchy metaphor

SWORDSwarm organizes agents in a corporate structure with 4 levels:

1. Executive Level (5 agents): DIRECTOR, CSO, LEADENGINEER, AGENTSMITH, PROJECTORCHESTRATOR
2. Senior Management (8 division heads): ARCHITECT, PLANNER, SECURITY, INFRASTRUCTURE, etc.
3. Middle Management: Division-specific managers
4. Specialists (many): PYTHON-INTERNAL, CPP-INTERNAL, ASSEMBLY-INTERNAL, APT41-DEFENSE, etc.

Technology bet

C/Rust binary layer + Python hooks + Intel NPU acceleration — an unusual focus on hardware performance for AI orchestration (typically GPU/CPU-bound by the LLM itself, not the orchestration layer).

Target users

(Claimed) enterprise teams needing 88 specialized agents with hardware acceleration and multi-IDE support.

SWORDSwarm — Architecture

Three-tier architecture

Tier 1: Binary Layer (C/Rust)
  - Hardware acceleration (Intel NPU, AVX2/AVX-512)
  - Performance-critical operations
  
Tier 2: Hook Layer (Python)
  - Python hook scripts
  - Interface between binary and agent layers
  - Orchestration logic
  
Tier 3: Agent Layer (Markdown)
  - 117 agent .md definition files
  - YAML frontmatter with metadata
  - Role descriptions and prompt content

Directory structure

SWORDSwarm/
├── agents/           # 117 agent .md files
├── orchestration/    # Orchestration scripts
│   ├── invoke.py     # Python subprocess bridge
│   ├── config.json
│   └── learning_system_tandem_orchestrator.py
├── hooks/            # Hook scripts
├── .claude/          # Claude Code integration
├── .cursor/          # Cursor IDE config
├── .windsurf/        # Windsurf config
├── .warp/            # Warp terminal config
├── .vscode/          # VS Code config
├── .mcp.json         # MCP configuration
├── config/           # System config
├── docs/             # Documentation
├── tests/            # Tests
├── benchmark/        # Performance benchmarks
├── security/         # Security configurations
├── crypto-pow/       # Crypto proof-of-work (unusual)
├── openvino/         # Intel OpenVINO integration
└── local-models/     # Local LLM model support

Required runtime

• Python 3.11+
• Virtual environment (venv)
• Intel NPU (optional, for hardware acceleration)
• OpenVINO (optional, for Intel acceleration)

Multi-IDE support

Separate config directories for: Claude Code (.claude), Cursor (.cursor), Windsurf (.windsurf), Warp (.warp), VS Code (.vscode) — enabling multi-IDE agent invocation.

SWORDSwarm — Components

Agent definitions (117 files in agents/)

Executive Level (5 agents)

Senior Management (8 agents)

ARCHITECT, PLANNER, SECURITY, INFRASTRUCTURE, DATASCIENCE, QADIRECTOR, ORCHESTRATOR, COORDINATOR

Selected specialists

PYTHON-INTERNAL, CPP-INTERNAL, C-INTERNAL, ASSEMBLY-INTERNAL, CRYPTO, CRYPTOEXPERT, AUDITOR, BASTION, AGENTSMITH, COGNITIVE_DEFENSE_AGENT, APT41-DEFENSE-AGENT, BGP-BLUE/RED/PURPLE-TEAM, CHAOS-AGENT, CISCO-AGENT, CLAUDECODE-PROMPTINJECTOR, CONSTRUCTOR, COORDINATOR, CARBON-INTERNAL-AGENT, C-MAKE-INTERNAL, CPP-GUI-INTERNAL, and 90+ more

Agent definition format (YAML+Markdown, v8.0 schema)

---
metadata:
  name: DIRECTOR
  version: 8.0.0
  uuid: d1r3c70r-3x3c-u71v-3000-57r4736y0001
  category: STRATEGIC
  priority: CRITICAL
  status: PRODUCTION
  color: "#FFD700"  # Gold - executive command authority
  emoji: "🎯"
  description: |
    ...multi-line description...
  tools:
    required:
      - Task  # MANDATORY for agent invocation
    code_operations: [Read, Write, Edit, MultiEdit]
    system_operations: [Bash, Grep, Glob, LS]
    information: [WebFetch, WebSearch]
    workflow: [TodoWrite, GitCommand]
  proactive_triggers:
    patterns: [...]
    context_triggers: [...]
    keywords: [...]
---
[Markdown role description content]

Orchestration scripts

Hardware acceleration

SWORDSwarm — Uniqueness & Positioning

differs_from_seeds

SWORDSwarm's agent definition format (YAML frontmatter + Markdown with UUID, version, proactive_triggers, tool declarations) is the most structured in this batch. The closest seed parallel is BMAD-METHOD's persona-md pattern (6 named persona files), but SWORDSwarm has 117 agents with explicit corporate hierarchy, UUIDs, color coding, and proactive trigger patterns. Unlike superpowers (14 skills, auto-triggered via SessionStart hook), SWORDSwarm's proactive triggers are specified in each agent's YAML and matched at inference time by Claude Code. The hardware acceleration (Intel NPU + OpenVINO) is unique in the entire corpus — no other framework attempts to accelerate the orchestration layer with specialized hardware. The multi-IDE config approach (.claude/, .cursor/, .windsurf/, .warp/, .vscode/) is also unique among frameworks studied.

Positioning (claimed)

• "Enterprise-grade" multi-agent orchestration
• "Military-grade optimization" (not independently verified)
• Hardware-first: Intel NPU as a differentiator
• Multi-IDE: not locked to Claude Code

Credibility concerns

Several signals suggest this is an experimental/hobbyist project despite enterprise claims:

• 24 stars (vs 172-5542 for other frameworks in this batch)
• Non-standard default branch (maine)
• No license file
• Performance metrics ("95% on-time delivery rate", "98.7% deployment success") are written into agent descriptions, not measured
• "I bolted warp on the side poorly" (README self-assessment)
• Last commit February 2026 (dormant)
• No public documentation site

Observable failure modes

• 117 agents × YAML+Markdown format = massive context injection risk
• No memory persistence: agents cannot learn or remember across sessions
• No git isolation: all agents operate in the same directory
• Corporate hierarchy may create bottlenecks (DIRECTOR as single point of failure)
• Hardware acceleration may be premature optimization (LLM inference dominates)

Inspired by

• Corporate org chart as agent hierarchy metaphor
• Intel NPU / OpenVINO ecosystem

Note on reliability

This report is based on accessible README and agent files only. Many claimed features (82% test coverage, CI/CD pipeline) could not be verified.