AgentArmy¶
AI-powered software development template with coordinated armies of specialized AI agents.
AgentArmy deploys two coordinated armies through a shared GitHub Projects v2 board:
- Claude Code army — local, deep, strategic: architecture, complex features, SAFE planning, security audits, data pipelines (via
dlt-engineerspecialist) - GitHub Copilot army — GitHub-native, fast, lightweight: PR review, simple task coding, board queries via
@board-manager
All significant work is tracked as GitHub issues on the project board. Agents operate as specialists — delegate to the right agent, in the right army, rather than doing everything generalist.
Quick Navigation¶
Getting Started
Set up your AgentArmy fork and configure GitHub Projects, Actions, and Copilot Chat.
Onboarding Check
Verify local auth, runner secrets, Project v2 access, and optional auto-add behavior before real work starts.
Agent Roster
Specialist agent definitions across 11 categories. Find the right expert for your task.
- :octocat: GitHub Copilot Army
Lightweight, GitHub-native automation for PR review, simple coding, and board queries.
GitHub Projects
Shared coordination plane for all armies. Issue types, fields, status workflow.
- :spiral_calendar: SAFe Framework
Program Increments (PIs), sprints, Definition of Ready/Done, velocity tracking.
Memory Palace
Organizational memory system for preserving context and learnings across agent sessions.
Human-in-the-Loop
Decision Artifacts surface creative and architectural forks to humans (or AI apps) via the board, without blocking parallel work.
Key Concepts¶
Two Armies¶
Each army has distinct strengths:
| Army | Strength | Use For |
|---|---|---|
| Claude Code | Deep thinking, complex architecture, multi-file refactors | Requirements analysis, system design, large features, security audits, data pipelines |
| GitHub Copilot | Fast turnaround, PR-native review, lightweight automation | Bug fixes, simple features (XS/S), inline PR feedback |
SAFE Planning¶
AgentArmy uses Scaled Agile Framework (SAFe) for predictable delivery:
- PI (Program Increment) = ~10 weeks of work
- Sprint = 2-week iterations
- Definition of Ready = Type, PI, Size, Estimate set + acceptance criteria
- Definition of Done = PR merged with
Closes #Nlink
GitHub Projects v2¶
Single source of truth for all work:
- Type field: Epic, Feature, Story, Enabler, Bug, Spike, Decision
- Status workflow: Todo → Ready → In Progress → In Review → Done (plus Awaiting Decision for HITL holds)
- PI field: Links issues to Program Increments
- Size & Estimate: T-shirt size + story points
Repository Structure¶
.claude/agents/categories/ → specialist agent definitions (11 categories)
.claude/commands/ → local slash commands (/wardley, /ea-adr, /capability-map)
.github/workflows/ → GitHub Actions: auto-status, routing, board commands, docs build
docs/ → This documentation
extensions/board-manager/ → Azure-deployable Copilot Chat extension
pipelines/ → dlt data pipeline toolkit
Getting Help¶
- Agent routing questions? See Agent Roster and Setup Guide
- GitHub Projects setup? See GitHub Projects Guide
- Runner token or setup validation? See Onboarding Sanity Check
- SAFE planning? See SAFe Framework
- Copilot Chat? See Copilot Army Guide