Skip to content

The Three Armies

AgentArmy coordinates three specialized armies, each with distinct strengths and operating models.

🧠 Claude Code Army

Depth, strategy, and architectural expertise.

Strengths

  • Deep thinking and complex problem-solving
  • Multi-file refactoring and architecture design
  • Security audits and compliance work
  • Enterprise patterns and scalability planning
  • Handles ambiguous requirements

How to Delegate

  1. Create an issue on the GitHub Projects board (hub repo only)
  2. Assign the correct specialist (e.g., backend-developer, react-specialist)
  3. Set Type, PI, and Size fields
  4. Agent picks up issue and works locally
  5. Opens PR with Closes #N link
  6. auto-status workflow moves issue to Done on merge

Note for spoke repos: Work arrives as Issues/PRs in your own repository, not on a hub board. The hub's scrum-master and product manager create these issues and assign them to you directly. See Agent Onboarding.

Best For

  • Requirements analysis and user story refinement
  • System architecture and design decisions
  • Large features (L/XL stories)
  • Code reviews and security audits
  • Complex refactoring

Response Time

Slower but deeper — typically hours to days depending on complexity.

⚡ GitHub Copilot Army

Speed, simplicity, and GitHub-native operations.

Strengths

  • Lightweight task automation
  • PR review and feedback
  • GitHub Projects board queries via @board-manager
  • Fast turnaround on simple tasks
  • Runs entirely within GitHub

How to Delegate

  1. Create an issue or label existing issue with copilot-task
  2. Copilot Coding Agent automatically:
  3. Creates a branch
  4. Implements the change
  5. Opens a PR with Closes #N
  6. Auto-merge on green CI

Best For

  • Bug fixes (XS/S size)
  • Simple features (S size, clear requirements)
  • PR reviews (automatic on all PRs)
  • Board queries and status checks
  • Label-based automation

Response Time

Fast — typically minutes to hours. Perfect for iteration velocity.

🚀 Antigravity CLI / Gemini Army

Autonomous execution with lightweight agentic control.

Strengths

  • Gemini's autonomous reasoning and planning
  • Lightweight agent execution (no heavy sessions)
  • CLI-native workflows and scripting
  • Quick research and knowledge synthesis
  • Multi-round problem solving with less overhead

How to Delegate

  1. Create an issue with specialized domain tags
  2. Trigger via Antigravity CLI: agents automatically synced from Claude definitions
  3. Or run ad-hoc: antigravity chat --agent security-architect

Best For

  • Quick research and exploration
  • CLI automation and scripting
  • Lightweight agents for exploratory work
  • Parallel multi-agent coordination
  • Autonomous problem decomposition

Response Time

Fast to very fast — optimized for lightweight execution and quick turnaround.

🔄 Shared Agent Definitions

All three armies read from the same agent definitions (.claude/agents/categories/):

  • Claude Code uses agents directly via Agent() tool
  • Codex receives synced agents in .codex/agents/ via SessionStart hook
  • Antigravity CLI receives agents organized as plugins in .agents/plugins/ (workspace) or ~/.gemini/antigravity-cli/plugins/ (global)

Synchronization mechanisms: - Codex: Automatic on SessionStart hook via .codex/hooks.json - Antigravity: Manual sync via orchestrator — python scripts/orchestrate_agent_sync.py --antigravity [--global] - GitHub apps: Receive routing definitions from source of truth

Edit agent definitions once in .claude/agents/categories/, then orchestrate synchronization across platforms.

See Agent Synchronization for full details and orchestration commands.

🗺️ Routing Matrix

Task Army Why
Fix a bug (XS/S) Copilot Fast, GitHub-native
Simple feature (S, clear spec) Copilot Speed and velocity
Large feature (L/XL) Claude Code Depth and architecture
Architect a system Claude Code Strategic thinking
API → database pipeline Claude Code (dlt-engineer) Pipeline specialists
Data warehouse prep Claude Code (dlt-engineer) ETL/ELT expertise
Security audit Claude Code Deep analysis
PR review (any size) Copilot Automatic + Copilot
Refactor monolith Claude Code Complex, multi-file
Simple code generation Copilot Speed
Quick research / exploration Antigravity CLI Lightweight, autonomous
Multi-agent exploration Antigravity CLI Parallel agent spawning
Ad-hoc scripting Antigravity CLI CLI-native workflows
Autonomous problem decomposition Antigravity CLI Gemini reasoning + lightweight execution

🎯 Principles

1. Route Early, Route Right

Avoid generalist work. Pick the specialist before creating the issue.

2. Define of Ready

Every issue needs: Type, PI, Size, Estimate, acceptance criteria.

3. Closes #N

Every PR must link back to its issue with Closes #N in the body.

4. Status Follows Work

auto-status workflow moves issues automatically. Don't manually update status unless correcting.

5. Specialize, Don't Duplicate

Armies don't overlap. Copilot handles GitHub-native work. Claude Code handles depth. dlt handles pipelines.

📚 Learn More