Agent Communication

Agent Communication

Skill for enabling real-time bidirectional communication between multiple AI agents working on the same project through a structured message board.

Input

• Agent identity (who is posting — e.g., "Manager", "UI Dev", "Backend Dev")
• Message type (status, decision, question, blocker)
• Message content

Output

• Shared board file at .agent/temp/board.md with structured messages
• Message history for audit trail and coordination

When to Use

Use this skill instead of agent-handoff when:

• ✅ Multiple agents are working simultaneously on the same codebase
• ✅ Agents need to coordinate decisions without blocking each other
• ✅ Manager Agent needs real-time visibility into worker progress
• ✅ Agents need to ask questions and get answers asynchronously

Use agent-handoff when:

• Agent A finishes → hands off complete work to Agent B (one-way delegation)

Process

Phase 1: Initialize Board

If .agent/temp/board.md doesn't exist, create it from the template:

1. Check if file exists: ls .agent/temp/board.md
2. If missing, copy template: use templates/BOARD_TEMPLATE.md
3. Replace {{PROJECT_NAME}} with actual project name
4. Replace {{DATE}} with current date

Phase 2: Post Message

When you have something to communicate:

1. Read current board state
2. Add your message under the appropriate section:
   • Status Updates → When you complete a task or start new work
   • Decision Log → When you make an architectural or design choice
   • Questions → When you need input from another agent
   • Blockers → When you can't proceed without something
3. Use this format:

**[Agent Name] — [Time]**
[Your message here]

Example:

**UI Dev — 2026-02-13 14:30**
✅ Completed login screen. Used MVVM pattern with StateFlow as discussed.

Phase 3: Read Board

Before starting work, check for relevant messages:

1. Read the entire board
2. Look for:
   • Questions addressed to you
   • Blockers that affect your work
   • Decisions that change your approach
3. Respond to questions by posting in the same section
4. Update status if you resolve a blocker

Phase 4: Archive (Optional)

When work is complete:

1. Move board to .agent/archive/board-[DATE].md (for historical reference)
2. Or delete if not needed for audit trail

Message Type Guide

Completeness Checklist

• □ Board file exists at .agent/temp/board.md?
• □ Messages are timestamped and attributed to an agent?
• □ All questions have been answered or acknowledged?
• □ No unresolved blockers remain?
• □ Decision log captures key architectural choices?

Rules

1. ALWAYS identify yourself when posting (use your agent role, e.g., "Manager", "UI Dev")
2. ALWAYS include a timestamp with your message
3. ALWAYS read the board before starting work to check for updates
4. ALWAYS respond to questions addressed to you
5. NEVER delete other agents' messages — only append
6. NEVER use the board for code — use PRs/commits for that
7. ALWAYS mark completed tasks with ✅ and blockers with 🚫
8. ALWAYS write in English for consistency
9. ALWAYS post a CLAIM message before modifying shared/foundational files (types, schemas, routes, configs). If another agent has already claimed the same file, negotiate ordering before proceeding.
10. NEVER start editing a file that another agent has an active CLAIM on without resolving the conflict first.

Example Board

# Agent Communication Board
**Project:** Order Management System
**Started:** 2026-02-13

---

## Status Updates

**Manager — 2026-02-13 10:00**
🚀 Project started. UI Dev: implement login. Backend Dev: set up database.

**UI Dev — 2026-02-13 14:30**
✅ Login screen complete. Used MVVM + StateFlow.

**Backend Dev — 2026-02-13 15:00**
✅ Database schema ready. Repository pattern implemented.

---

## Decision Log

**Manager — 2026-02-13 10:15**
Decision: Use Room for local storage (offline-first requirement).

**UI Dev — 2026-02-13 12:00**
Decision: Dark mode will be default. Light mode in phase 2.

---

## Questions

**UI Dev — 2026-02-13 11:00**
Q for Backend Dev: What's the error format for failed logins?

**Backend Dev — 2026-02-13 11:30**
A: Use `Result<User, AuthError>` wrapper. See `models/Result.kt`.

---

## Blockers

**Backend Dev — 2026-02-13 16:00**
🚫 Blocked: Need Firebase API key to test authentication.

**Manager — 2026-02-13 16:15**
✅ Unblocked: API key sent via Slack. Check `.env` file.