ELF Swarm Coordination

Orchestrate multi-agent workflows using the full agent pool (~100+ specialized agents).

Swarm Modes

ultrathink

Maximum depth analysis. Launch 10-20+ agents across all relevant categories.

/swarm ultrathink [target]

focused

Targeted analysis. Launch 3-5 agents for specific domain.

/swarm focused security [target]
/swarm focused architecture [target]

quick

Fast survey. Launch 2-3 core agents.

/swarm quick [target]

Agent Selection Logic

DO NOT hardcode agents. Select based on task characteristics:

Step 1: Detect Task Domains

Analyze the request and codebase to identify:

• Languages present (Python, TypeScript, Rust, Go, etc.)
• Frameworks (React, FastAPI, Django, etc.)
• Infrastructure (Docker, K8s, Terraform, etc.)
• Concerns (security, performance, architecture, etc.)

Step 2: Map Domains to Agent Categories

Domain	Agents to Consider
Code Quality	code-reviewer, debugger, test-automator
Architecture	architect-review, backend-architect, cloud-architect, database-architect
Security	security-auditor, backend-security-coder, frontend-security-coder, mobile-security-coder
Python	python-pro, fastapi-pro, django-pro, python-testing-patterns
TypeScript/JS	typescript-pro, javascript-pro, frontend-developer, react-state-management
Rust	rust-pro, rust-async-patterns, memory-safety-patterns
Go	golang-pro, go-concurrency-patterns
Databases	database-architect, database-optimizer, database-admin, sql-pro
Infrastructure	devops-troubleshooter, kubernetes-architect, terraform-specialist, deployment-engineer
Documentation	docs-architect, tutorial-engineer, reference-builder, api-documenter
Performance	performance-engineer, database-optimizer
AI/Agents	ai-engineer, prompt-engineer, context-manager
Frontend	frontend-developer, ui-ux-designer, tailwind-design-system
DevEx	dx-optimizer
Testing	test-automator, tdd-orchestrator, e2e-testing-patterns
Shell/Scripts	bash-pro, posix-shell-pro, shellcheck-configuration

Step 3: Select Agent Count by Mode

Mode	Agents per Category	Total Target
ultrathink	2-3	15-25
focused	1-2	4-8
quick	1	2-4

Execution Rules

1. Always async: run_in_background=True for ALL agents
2. Parallel launch: Send ALL agent spawns in ONE message
3. Block only at end: Use TaskOutput only when aggregating results
4. Model selection:
   • Haiku for quick/simple analysis
   • Sonnet for standard analysis (default)
   • Opus for deep architectural/security audits

Example: ultrathink on a Python/React Project

Detected: Python backend, React frontend, SQLite database, shell scripts

Agents to launch:

# Code Quality
- code-reviewer
- debugger

# Architecture
- architect-review
- backend-architect
- database-architect

# Security
- security-auditor
- backend-security-coder
- frontend-security-coder

# Language-Specific
- python-pro
- typescript-pro
- frontend-developer

# Database
- database-optimizer

# Documentation
- docs-architect

# DevEx
- dx-optimizer

# Testing
- test-automator

# Shell
- bash-pro

# AI (if agent framework)
- prompt-engineer
- context-manager
- ai-engineer

Total: 18 agents in parallel

Prompt Template for Agents

Each agent gets a focused prompt:

[Agent Type] analysis of [TARGET_PATH].

Focus on:
- [Domain-specific concerns]
- [What to look for]
- [What to report]

Be thorough. Report findings with file:line references.

Result Aggregation

After all agents complete:

1. Read all output files
2. Group findings by severity/category
3. Identify patterns across agents (multiple agents flagging same issue = high confidence)
4. Synthesize into actionable summary
5. Optionally record learnings to ELF building

Anti-Patterns (DO NOT DO)

• Hardcoding 4 agents (Researcher/Architect/Creative/Skeptic is OBSOLETE)
• Launching agents synchronously
• Using same prompt for all agents
• Ignoring detected technologies
• Using Opus for everything (wasteful)

Integration with ELF

After swarm completes:

• Record significant findings as heuristics
• Update golden rules if patterns emerge
• Escalate architectural decisions to CEO inbox