Agent Communication Skill

Description

Comprehensive inter-agent messaging and coordination system providing protocol-based communication, shared context management, event broadcasting with subscriptions, automated handoff documentation, and message persistence. This skill enables agents to coordinate work, share decisions, negotiate contracts, and maintain consistent state across distributed parallel execution environments.

When to Use This Skill

"Facilitate communication between builder agents needing to coordinate API contract changes"
"Coordinate handoffs between architect completing design and builders starting implementation"
"Manage inter-agent messaging for microservices team sharing service interface definitions"
"Broadcast architectural decisions to all agents working on authentication feature"
"Handle agent-to-agent communication for parallel data pipeline development with shared schemas"

When NOT to Use This Skill

Single agent workflows with no coordination needs → No messaging required
Sequential handoffs with explicit synchronization → Use standard handoff documentation
Simple file-based coordination (shared README) → Direct file collaboration sufficient
User-to-agent communication → Use standard CLI/UI interactions

Prerequisites

Multiple agents working concurrently or in sequence
Shared filesystem or message bus infrastructure
Agent registry with contact information and capabilities
Standardized message schema and protocols
Event subscription system (file-based or message queue)

Workflow

Phase 1: Communication Infrastructure Setup

Purpose: Establish message transport, routing, and persistence mechanisms

Step 1.1: Initialize Message Bus

Set up message routing infrastructure:

bash

function initialize_message_bus() {
  local project_root=$1

  echo "Initializing agent message bus..."

  # Create message directories
  mkdir -p "$project_root/.agent-messages/inbox"
  mkdir -p "$project_root/.agent-messages/outbox"
  mkdir -p "$project_root/.agent-messages/archive"
  mkdir -p "$project_root/.agent-messages/events"
  mkdir -p "$project_root/.agent-messages/broadcasts"

  # Create message registry
  cat > "$project_root/.agent-messages/registry.json" <<EOF
{
  "initialized_at": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
  "message_count": 0,
  "agents": {},
  "event_subscriptions": {},
  "message_retention_hours": 72
}
EOF

  # Create routing table
  cat > "$project_root/.agent-messages/routing.json" <<EOF
{
  "routes": {
    "direct": "inbox/{recipient}/",
    "broadcast": "broadcasts/",
    "event": "events/{event_type}/",
    "archive": "archive/{date}/"
  },
  "priority_queues": {
    "critical": 1,
    "high": 2,
    "normal": 3,
    "low": 4
  }
}
EOF

  echo "✅ Message bus initialized at $project_root/.agent-messages/"
}

Directory Structure:

.agent-messages/
├── inbox/              # Per-agent inboxes
│   ├── builder-1/
│   ├── builder-2/
│   └── validator-1/
├── outbox/             # Pending message delivery
├── archive/            # Historical messages (by date)
│   ├── 2025-11-29/
│   └── 2025-11-30/
├── events/             # Event-based messages
│   ├── TaskCompleted/
│   ├── BlockerEncountered/
│   └── MilestoneReached/
├── broadcasts/         # Messages to all agents
└── registry.json       # Message bus metadata

Expected Output: Initialized message bus directory structure with routing configuration

Step 1.2: Register Agents

Build agent directory for routing:

bash

function register_agent() {
  local agent_id=$1
  local agent_role=$2
  local worktree_path=$3
  local capabilities=$4  # JSON array

  local registry_file=".agent-messages/registry.json"

  # Create agent inbox
  mkdir -p ".agent-messages/inbox/$agent_id"

  # Register agent in directory
  jq --arg id "$agent_id" \
     --arg role "$agent_role" \
     --arg path "$worktree_path" \
     --argjson caps "$capabilities" \
     '.agents[$id] = {
       "role": $role,
       "worktree_path": $path,
       "capabilities": $caps,
       "registered_at": now|todate,
       "status": "active",
       "last_seen": now|todate,
       "message_count": 0
     }' "$registry_file" > "$registry_file.tmp"

  mv "$registry_file.tmp" "$registry_file"

  echo "✅ Registered agent: $agent_id ($agent_role)"
}

# Example registration
register_agent "builder-1" "builder" "worktrees/builder-1" \
  '["jwt-implementation", "api-development", "typescript"]'

Agent Registry Entry:

json

{
  "agents": {
    "builder-1": {
      "role": "builder",
      "worktree_path": "worktrees/builder-1",
      "capabilities": ["jwt-implementation", "api-development", "typescript"],
      "registered_at": "2025-11-29T14:30:52Z",
      "status": "active",
      "last_seen": "2025-11-29T14:30:52Z",
      "message_count": 0
    }
  }
}

Expected Output: Registered agents with inboxes and capability metadata

Step 1.3: Define Message Schema

Establish standardized message format:

Message Schema (JSON):

json

{
  "message_id": "msg_20251129_143052_001",
  "protocol_version": "1.0",
  "timestamp": "2025-11-29T14:30:52Z",
  "from_agent": "builder-1",
  "to_agent": "builder-2",
  "message_type": "interface_contract",
  "priority": "high",
  "subject": "IUserService interface definition",
  "payload": {
    "contract_type": "typescript_interface",
    "interface_name": "IUserService",
    "methods": [
      {
        "name": "findById",
        "params": [{"name": "id", "type": "string"}],
        "return_type": "Promise<User | null>"
      },
      {
        "name": "create",
        "params": [{"name": "userData", "type": "CreateUserDto"}],
        "return_type": "Promise<User>"
      }
    ],
    "purpose": "JWT service needs to lookup users by ID for token generation"
  },
  "requires_response": true,
  "response_deadline": "2025-11-29T16:30:52Z",
  "conversation_id": "conv_auth_contracts_001",
  "metadata": {
    "task_id": "pg2-task-1",
    "related_files": ["services/user.service.ts", "services/jwt.service.ts"]
  }
}

Message Types:

interface_contract: API/interface definitions
decision_announcement: Architectural decisions
blocker_notification: Work blocked, needs input
handoff_request: Ready to transfer work
resource_conflict: File/resource contention
progress_update: Status notification
question: Request for clarification
response: Reply to previous message

Expected Output: Message schema documentation and validation library

Phase 2: Message Sending and Routing

Purpose: Enable agents to send messages with appropriate routing and delivery

Step 2.1: Send Direct Message

Agent-to-agent point-to-point messaging:

bash

function send_message() {
  local from_agent=$1
  local to_agent=$2
  local message_type=$3
  local subject=$4
  local payload=$5  # JSON string
  local priority=${6:-normal}

  local message_id="msg_$(date +%Y%m%d_%H%M%S)_$(random_string 6)"
  local timestamp=$(date -u +%Y-%m-%dT%H:%M:%SZ)

  # Create message file
  local message_file=".agent-messages/inbox/$to_agent/${message_id}.json"

  cat > "$message_file" <<EOF
{
  "message_id": "$message_id",
  "protocol_version": "1.0",
  "timestamp": "$timestamp",
  "from_agent": "$from_agent",
  "to_agent": "$to_agent",
  "message_type": "$message_type",
  "priority": "$priority",
  "subject": "$subject",
  "payload": $payload,
  "requires_response": false,
  "status": "delivered",
  "delivered_at": "$timestamp"
}
EOF

  # Update message count
  jq --arg agent "$to_agent" \
     '.agents[$agent].message_count += 1' \
     .agent-messages/registry.json > .agent-messages/registry.json.tmp
  mv .agent-messages/registry.json.tmp .agent-messages/registry.json

  # Log to outbox for sender
  cp "$message_file" ".agent-messages/outbox/${message_id}_to_${to_agent}.json"

  echo "✅ Message sent: $from_agent → $to_agent ($message_type)"
  echo "   ID: $message_id"
  echo "   Subject: $subject"
}

# Example usage
send_message "builder-1" "builder-2" "interface_contract" \
  "IUserService interface definition" \
  '{"interface": "IUserService", "methods": ["findById", "create"]}' \
  "high"

Expected Output: Message delivered to recipient inbox with confirmation

Step 2.2: Broadcast Message

Send message to all registered agents:

bash

function broadcast_message() {
  local from_agent=$1
  local message_type=$2
  local subject=$3
  local payload=$4

  local message_id="bcast_$(date +%Y%m%d_%H%M%S)_$(random_string 6)"
  local timestamp=$(date -u +%Y-%m-%dT%H:%M:%SZ)

  # Create broadcast message
  local broadcast_file=".agent-messages/broadcasts/${message_id}.json"

  cat > "$broadcast_file" <<EOF
{
  "message_id": "$message_id",
  "protocol_version": "1.0",
  "timestamp": "$timestamp",
  "from_agent": "$from_agent",
  "to_agent": "ALL",
  "message_type": "$message_type",
  "subject": "$subject",
  "payload": $payload,
  "broadcast": true
}
EOF

  # Deliver to all active agents
  local agent_count=0
  for agent_id in $(jq -r '.agents | keys[]' .agent-messages/registry.json); do
    if [ "$agent_id" != "$from_agent" ]; then
      cp "$broadcast_file" ".agent-messages/inbox/$agent_id/${message_id}.json"
      agent_count=$((agent_count + 1))
    fi
  done

  echo "✅ Broadcast sent to $agent_count agents"
  echo "   From: $from_agent"
  echo "   Subject: $subject"
}

# Example broadcast
broadcast_message "architect-1" "decision_announcement" \
  "Authentication will use RS256 algorithm for JWT" \
  '{
    "decision": "Use RS256 for JWT signing",
    "rationale": "Better security, supports key rotation",
    "impact": "All auth services must use RS256 keypair",
    "effective_date": "immediate"
  }'

Expected Output: Message delivered to all agent inboxes

Step 2.3: Publish Event

Emit event for subscribed agents:

bash

function publish_event() {
  local event_type=$1
  local source_agent=$2
  local event_data=$3  # JSON

  local event_id="evt_$(date +%Y%m%d_%H%M%S)_$(random_string 6)"
  local timestamp=$(date -u +%Y-%m-%dT%H:%M:%SZ)

  # Create event directory if needed
  mkdir -p ".agent-messages/events/$event_type"

  # Create event file
  local event_file=".agent-messages/events/$event_type/${event_id}.json"

  cat > "$event_file" <<EOF
{
  "event_id": "$event_id",
  "event_type": "$event_type",
  "timestamp": "$timestamp",
  "source_agent": "$source_agent",
  "event_data": $event_data
}
EOF

  # Deliver to subscribed agents
  local subscribers=$(jq -r ".event_subscriptions[\"$event_type\"] // []" \
    .agent-messages/registry.json)

  for agent_id in $(echo "$subscribers" | jq -r '.[]'); do
    cp "$event_file" ".agent-messages/inbox/$agent_id/${event_id}.json"
  done

  echo "✅ Event published: $event_type"
  echo "   Source: $source_agent"
  echo "   Subscribers notified: $(echo "$subscribers" | jq length)"
}

# Example event
publish_event "TaskCompleted" "builder-1" \
  '{
    "task_id": "pg2-task-1",
    "task_name": "Implement JWT service",
    "completed_at": "2025-11-29T16:15:30Z",
    "artifacts": ["services/jwt.service.ts", "services/jwt.service.spec.ts"],
    "test_results": {"passed": 24, "failed": 0, "coverage": 92}
  }'

Expected Output: Event delivered to all subscribed agents

Phase 3: Message Receiving and Processing

Purpose: Enable agents to poll, read, and process incoming messages

Step 3.1: Check Inbox for New Messages

Poll for unread messages:

bash

function check_inbox() {
  local agent_id=$1
  local inbox_dir=".agent-messages/inbox/$agent_id"

  # Find unread messages (no .read marker)
  local unread_messages=()

  for msg_file in "$inbox_dir"/*.json; do
    if [ -f "$msg_file" ] && [ ! -f "${msg_file}.read" ]; then
      unread_messages+=("$msg_file")
    fi
  done

  echo "📬 Inbox check for $agent_id"
  echo "   Unread messages: ${#unread_messages[@]}"

  # Display message summary
  for msg_file in "${unread_messages[@]}"; do
    local from=$(jq -r '.from_agent' "$msg_file")
    local type=$(jq -r '.message_type' "$msg_file")
    local subject=$(jq -r '.subject' "$msg_file")
    local priority=$(jq -r '.priority' "$msg_file")

    echo "   [$priority] $from: $subject ($type)"
  done

  # Return unread message files
  printf '%s\n' "${unread_messages[@]}"
}

Expected Output: List of unread messages with metadata

Step 3.2: Read and Process Message

Retrieve and handle specific message:

bash

function read_message() {
  local agent_id=$1
  local message_file=$2

  echo "Reading message: $(basename "$message_file")"

  # Parse message
  local from_agent=$(jq -r '.from_agent' "$message_file")
  local message_type=$(jq -r '.message_type' "$message_file")
  local subject=$(jq -r '.subject' "$message_file")
  local payload=$(jq -r '.payload' "$message_file")
  local requires_response=$(jq -r '.requires_response' "$message_file")

  echo "From: $from_agent"
  echo "Type: $message_type"
  echo "Subject: $subject"
  echo ""
  echo "Payload:"
  echo "$payload" | jq .

  # Process based on message type
  case "$message_type" in
    interface_contract)
      process_interface_contract "$agent_id" "$message_file"
      ;;
    decision_announcement)
      process_decision "$agent_id" "$message_file"
      ;;
    blocker_notification)
      process_blocker "$agent_id" "$message_file"
      ;;
    handoff_request)
      process_handoff "$agent_id" "$message_file"
      ;;
    question)
      process_question "$agent_id" "$message_file"
      ;;
    *)
      echo "⚠️  Unknown message type: $message_type"
      ;;
  esac

  # Mark as read
  touch "${message_file}.read"
  echo "$(date -u +%Y-%m-%dT%H:%M:%SZ)" > "${message_file}.read"

  # Send acknowledgment if required
  if [ "$requires_response" == "true" ]; then
    send_acknowledgment "$agent_id" "$from_agent" "$message_file"
  fi
}

Expected Output: Processed message with appropriate action taken

Step 3.3: Respond to Message

Send reply to received message:

bash

function respond_to_message() {
  local agent_id=$1
  local original_message_file=$2
  local response_payload=$3  # JSON

  local original_id=$(jq -r '.message_id' "$original_message_file")
  local original_from=$(jq -r '.from_agent' "$original_message_file")
  local conversation_id=$(jq -r '.conversation_id // empty' "$original_message_file")

  # If no conversation ID, create one
  if [ -z "$conversation_id" ]; then
    conversation_id="conv_$(date +%Y%m%d_%H%M%S)"
  fi

  local response_id="resp_$(date +%Y%m%d_%H%M%S)_$(random_string 6)"
  local timestamp=$(date -u +%Y-%m-%dT%H:%M:%SZ)

  # Create response message
  local response_file=".agent-messages/inbox/$original_from/${response_id}.json"

  cat > "$response_file" <<EOF
{
  "message_id": "$response_id",
  "protocol_version": "1.0",
  "timestamp": "$timestamp",
  "from_agent": "$agent_id",
  "to_agent": "$original_from",
  "message_type": "response",
  "subject": "Re: $(jq -r '.subject' "$original_message_file")",
  "payload": $response_payload,
  "in_reply_to": "$original_id",
  "conversation_id": "$conversation_id",
  "status": "delivered"
}
EOF

  echo "✅ Response sent to $original_from"
  echo "   Conversation: $conversation_id"
}

# Example response
respond_to_message "builder-2" "$message_file" \
  '{
    "status": "accepted",
    "interface_approved": true,
    "implementation_notes": "Will implement IUserService with these methods",
    "estimated_completion": "2025-11-29T18:00:00Z"
  }'

Expected Output: Response message delivered to original sender

Phase 4: Shared Context Management

Purpose: Maintain synchronized state and shared knowledge across agents

Step 4.1: Create Shared Context Document

Establish shared knowledge base:

bash

function initialize_shared_context() {
  local context_name=$1
  local initial_content=$2

  mkdir -p ".agent-messages/shared-context"

  local context_file=".agent-messages/shared-context/${context_name}.md"

  cat > "$context_file" <<EOF
# Shared Context: $context_name

**Created**: $(date -u +%Y-%m-%dT%H:%M:%SZ)
**Last Updated**: $(date -u +%Y-%m-%dT%H:%M:%SZ)
**Contributors**: []

---

$initial_content

---

## Update History
- $(date -u +%Y-%m-%dT%H:%M:%S)Z: Initial creation
EOF

  # Create lock file for concurrent access control
  cat > "$context_file.lock" <<EOF
{
  "locked": false,
  "locked_by": null,
  "locked_at": null
}
EOF

  echo "✅ Created shared context: $context_name"
}

# Example: Create API contracts context
initialize_shared_context "api-contracts" \
  "# API Contract Definitions

## Authentication Endpoints

### POST /auth/login
**Status**: Defined by architect-1
**Implemented by**: builder-1 (in progress)

### POST /auth/register
**Status**: Pending definition
"

Shared Context Types:

API Contracts: Interface definitions, endpoint specs
Architecture Decisions: Design choices with rationale
Data Models: Database schemas, entity relationships
Configuration Standards: Naming conventions, standards
Integration Points: Service boundaries, dependencies

Expected Output: Initialized shared context document with version control

Step 4.2: Update Shared Context

Concurrent-safe context updates:

bash

function update_shared_context() {
  local context_name=$1
  local agent_id=$2
  local update_content=$3

  local context_file=".agent-messages/shared-context/${context_name}.md"
  local lock_file="${context_file}.lock"

  # Acquire lock
  local max_wait=30  # seconds
  local waited=0

  while [ $waited -lt $max_wait ]; do
    local is_locked=$(jq -r '.locked' "$lock_file")

    if [ "$is_locked" == "false" ]; then
      # Acquire lock
      jq --arg agent "$agent_id" \
         '.locked = true | .locked_by = $agent | .locked_at = now|todate' \
         "$lock_file" > "${lock_file}.tmp"
      mv "${lock_file}.tmp" "$lock_file"
      break
    fi

    sleep 1
    waited=$((waited + 1))
  done

  if [ $waited -eq $max_wait ]; then
    echo "❌ Failed to acquire lock on $context_name after ${max_wait}s"
    return 1
  fi

  # Update context
  {
    cat "$context_file"
    echo ""
    echo "## Update: $(date -u +%Y-%m-%dT%H:%M:%S)Z by $agent_id"
    echo "$update_content"
  } > "${context_file}.tmp"

  mv "${context_file}.tmp" "$context_file"

  # Update metadata
  sed -i.bak "s/Last Updated:.*/Last Updated: $(date -u +%Y-%m-%dT%H:%M:%SZ)/" "$context_file"

  # Release lock
  jq '.locked = false | .locked_by = null | .locked_at = null' \
     "$lock_file" > "${lock_file}.tmp"
  mv "${lock_file}.tmp" "$lock_file"

  echo "✅ Updated shared context: $context_name"

  # Broadcast update notification
  broadcast_message "$agent_id" "context_updated" \
    "Shared context '$context_name' updated" \
    "{\"context\": \"$context_name\", \"updated_by\": \"$agent_id\"}"
}

# Example update
update_shared_context "api-contracts" "builder-1" \
  "### POST /auth/login
**Implementation complete**
- JWT token returned in response
- Refresh token stored in httpOnly cookie
- 2FA support included
"

Expected Output: Updated shared context with change notification broadcast

Phase 5: Handoff Documentation

Purpose: Formalize work transitions between agents with complete context

Step 5.1: Create Handoff Package

Bundle all artifacts and context for receiving agent:

bash

function create_handoff_package() {
  local from_agent=$1
  local to_agent=$2
  local task_description=$3
  local artifacts=$4  # JSON array of file paths
  local context=$5    # Additional context text

  local handoff_id="handoff_$(date +%Y%m%d_%H%M%S)"
  local handoff_dir=".agent-messages/handoffs/$handoff_id"

  mkdir -p "$handoff_dir/artifacts"

  # Create handoff document
  cat > "$handoff_dir/HANDOFF.md" <<EOF
# Handoff: $from_agent → $to_agent

**Handoff ID**: $handoff_id
**Created**: $(date -u +%Y-%m-%dT%H:%M:%SZ)
**From**: $from_agent
**To**: $to_agent

## Task Description
$task_description

## Work Completed
[Summary of completed work]

## Artifacts Transferred
$(echo "$artifacts" | jq -r '.[] | "- " + .')

## Context and Background
$context

## Next Steps for Receiving Agent
1. Review artifacts in \`$handoff_dir/artifacts/\`
2. Read shared context documents
3. Run tests to verify functionality
4. Continue implementation from this point

## Known Issues / Blockers
[Any issues or blockers to be aware of]

## Questions / Clarifications
[Areas needing clarification]

## Acceptance Criteria
- [ ] All tests passing
- [ ] Code reviewed and understood
- [ ] Integration points verified
- [ ] Ready to continue implementation
EOF

  # Copy artifacts
  for artifact in $(echo "$artifacts" | jq -r '.[]'); do
    cp "$artifact" "$handoff_dir/artifacts/"
  done

  # Create metadata
  cat > "$handoff_dir/metadata.json" <<EOF
{
  "handoff_id": "$handoff_id",
  "from_agent": "$from_agent",
  "to_agent": "$to_agent",
  "created_at": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
  "status": "pending_acceptance",
  "artifacts": $artifacts
}
EOF

  # Send handoff notification
  send_message "$from_agent" "$to_agent" "handoff_request" \
    "Handoff ready: $task_description" \
    "{\"handoff_id\": \"$handoff_id\", \"handoff_dir\": \"$handoff_dir\"}" \
    "high"

  echo "✅ Handoff package created: $handoff_id"
  echo "   From: $from_agent → $to_agent"
  echo "   Artifacts: $(echo "$artifacts" | jq length)"
}

# Example handoff
create_handoff_package "architect-1" "builder-1" \
  "Implement JWT service based on design" \
  '["docs/jwt-design.md", "interfaces/IJwtService.ts", "schemas/token-schema.json"]' \
  "JWT service design is complete. Using RS256 algorithm. See shared context for API contracts."

Handoff Package Contents:

HANDOFF.md: Complete handoff documentation
artifacts/: All relevant files and documents
metadata.json: Structured handoff metadata
context-links.md: References to shared context docs
acceptance-checklist.md: Criteria for accepting handoff

Expected Output: Complete handoff package with notification sent

Step 5.2: Accept Handoff

Receiving agent accepts and acknowledges handoff:

bash

function accept_handoff() {
  local agent_id=$1
  local handoff_id=$2
  local acceptance_notes=$3

  local handoff_dir=".agent-messages/handoffs/$handoff_id"
  local metadata_file="$handoff_dir/metadata.json"

  # Update handoff status
  jq --arg agent "$agent_id" \
     --arg notes "$acceptance_notes" \
     '.status = "accepted" |
      .accepted_by = $agent |
      .accepted_at = now|todate |
      .acceptance_notes = $notes' \
     "$metadata_file" > "${metadata_file}.tmp"
  mv "${metadata_file}.tmp" "$metadata_file"

  # Append to handoff document
  cat >> "$handoff_dir/HANDOFF.md" <<EOF

---

## Handoff Accepted

**Accepted by**: $agent_id
**Accepted at**: $(date -u +%Y-%m-%dT%H:%M:%SZ)

### Acceptance Notes
$acceptance_notes

### Acceptance Checklist
- [x] All tests passing
- [x] Code reviewed and understood
- [x] Integration points verified
- [x] Ready to continue implementation
EOF

  # Notify original agent
  local from_agent=$(jq -r '.from_agent' "$metadata_file")
  send_message "$agent_id" "$from_agent" "handoff_accepted" \
    "Handoff $handoff_id accepted" \
    "{\"handoff_id\": \"$handoff_id\", \"notes\": \"$acceptance_notes\"}"

  echo "✅ Handoff accepted: $handoff_id"
}

# Example acceptance
accept_handoff "builder-1" "handoff_20251129_143052" \
  "Design reviewed. All acceptance criteria met. Starting implementation."

Expected Output: Updated handoff status with acceptance confirmation

Examples

Example 1: Interface Contract Negotiation

Context: Two builder agents need to coordinate API interface changes

Execution Flow:

bash

# Builder-1 proposes interface
send_message "builder-1" "builder-2" "interface_contract" \
  "Proposed IUserService interface" \
  '{
    "interface": "IUserService",
    "methods": [
      {"name": "findById", "params": ["id: string"], "returns": "Promise<User>"},
      {"name": "create", "params": ["data: CreateUserDto"], "returns": "Promise<User>"}
    ]
  }' "high"

# Builder-2 receives and reviews
check_inbox "builder-2"
# Output: 1 unread message from builder-1

read_message "builder-2" ".agent-messages/inbox/builder-2/msg_20251129_143052_abc123.json"

# Builder-2 proposes modification
respond_to_message "builder-2" "$message_file" \
  '{
    "status": "modification_requested",
    "changes": [
      {
        "method": "findById",
        "change": "Return type should be Promise<User | null> to handle not found"
      }
    ],
    "rationale": "Avoids throwing exceptions for normal control flow"
  }'

# Builder-1 accepts modification
respond_to_message "builder-1" "$response_file" \
  '{
    "status": "accepted",
    "updated_interface": {
      "interface": "IUserService",
      "methods": [
        {"name": "findById", "params": ["id: string"], "returns": "Promise<User | null>"},
        {"name": "create", "params": ["data: CreateUserDto"], "returns": "Promise<User>"}
      ]
    }
  }'

# Update shared context with agreed interface
update_shared_context "api-contracts" "builder-1" \
  "### IUserService Interface (AGREED)
\`\`\`typescript
interface IUserService {
  findById(id: string): Promise<User | null>;
  create(data: CreateUserDto): Promise<User>;
}
\`\`\`
**Agreed by**: builder-1, builder-2
**Date**: 2025-11-29
"

Expected Output:

Conversation Summary:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Conversation ID: conv_auth_contracts_001
Participants: builder-1, builder-2
Messages: 3
Duration: 8 minutes
Status: ✅ RESOLVED

Outcome:
- IUserService interface defined
- Method signature modified (findById returns nullable)
- Agreement documented in shared context
- Both agents ready to implement

Rationale: Demonstrates collaborative contract negotiation with proposal, feedback, and agreement

Example 2: Architectural Decision Broadcast

Context: Architect makes decision affecting all agents

Execution Flow:

bash

# Architect broadcasts decision
broadcast_message "architect-1" "decision_announcement" \
  "JWT Algorithm Decision: RS256" \
  '{
    "decision_id": "DEC-001",
    "decision": "Use RS256 algorithm for JWT signing",
    "rationale": [
      "Better security than HS256",
      "Supports key rotation",
      "Industry best practice for production systems"
    ],
    "impact": {
      "affected_components": ["jwt-service", "auth-middleware", "token-validator"],
      "breaking_change": false,
      "action_required": "All JWT implementations must use RS256 keypair"
    },
    "effective_date": "immediate",
    "references": [
      "https://tools.ietf.org/html/rfc7519",
      "docs/security/jwt-best-practices.md"
    ]
  }'

# All agents receive and acknowledge
for agent in builder-1 builder-2 validator-1; do
  # Agents check inbox
  unread=$(check_inbox "$agent")

  # Agents read decision
  read_message "$agent" "$unread"

  # Agents acknowledge
  send_message "$agent" "architect-1" "acknowledgment" \
    "Acknowledged decision DEC-001" \
    '{"decision_id": "DEC-001", "will_comply": true}'
done

# Update shared context
update_shared_context "architecture-decisions" "architect-1" \
  "## DEC-001: JWT Algorithm Selection

**Decision**: Use RS256 algorithm for JWT signing
**Date**: 2025-11-29
**Status**: APPROVED

**Rationale**:
- Better security than HS256
- Supports key rotation
- Industry best practice

**Impact**: All JWT services must use RS256 keypair
**Acknowledgments**: builder-1 ✅, builder-2 ✅, validator-1 ✅
"

Expected Output:

Broadcast Report:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Decision: DEC-001 (JWT Algorithm: RS256)
Broadcast at: 2025-11-29T14:30:52Z
Recipients: 3 agents

Acknowledgments:
✅ builder-1 (received: 14:31, acknowledged: 14:32)
✅ builder-2 (received: 14:31, acknowledged: 14:33)
✅ validator-1 (received: 14:31, acknowledged: 14:35)

Status: Fully acknowledged (3/3)
Shared context updated: architecture-decisions.md

Rationale: Shows broadcast pattern for architectural decisions with acknowledgment tracking

Example 3: Blocker Escalation and Resolution

Context: Agent encounters blocker needing orchestrator intervention

Execution Flow:

bash

# Builder-1 hits blocker
publish_event "BlockerEncountered" "builder-1" \
  '{
    "task_id": "pg2-task-1",
    "blocker_type": "missing_dependency",
    "description": "Cannot implement JWT refresh without User model",
    "blocked_since": "2025-11-29T15:30:00Z",
    "dependency": {
      "required_artifact": "User model with refreshToken field",
      "responsible_agent": "builder-2",
      "estimated_impact": "2 hours blocked"
    }
  }'

# Orchestrator receives event (subscribed to BlockerEncountered)
# Checks builder-2 status
builder_2_status=$(jq -r '.status' "worktrees/builder-2/status/agent-status.json")

# Orchestrator facilitates communication
send_message "orchestrator" "builder-2" "priority_request" \
  "Builder-1 blocked waiting for User model" \
  '{
    "urgency": "high",
    "blocked_agent": "builder-1",
    "required_artifact": "User model with refreshToken field",
    "current_eta": "unknown",
    "requested_eta": "within 30 minutes"
  }'

# Builder-2 responds with ETA
respond_to_message "builder-2" "$message_file" \
  '{
    "status": "acknowledged",
    "current_progress": "User model 75% complete",
    "eta": "2025-11-29T16:15:00Z",
    "partial_delivery": "Can provide interface definition immediately"
  }'

# Builder-2 sends partial interface to unblock builder-1
send_message "builder-2" "builder-1" "interface_contract" \
  "Partial User interface (full implementation coming soon)" \
  '{
    "interface": "User",
    "fields": ["id", "email", "refreshToken"],
    "note": "Full implementation ETA 16:15, use this interface now"
  }'

# Builder-1 unblocked
publish_event "BlockerResolved" "builder-1" \
  '{
    "task_id": "pg2-task-1",
    "blocker_type": "missing_dependency",
    "resolution": "Partial interface received from builder-2",
    "blocked_duration": "25 minutes",
    "resolved_at": "2025-11-29T15:55:00Z"
  }'

Expected Output:

Blocker Resolution Report:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Blocker ID: blocker_pg2_task1_001
Reported by: builder-1
Blocker type: missing_dependency

Timeline:
15:30:00 - Blocker encountered
15:32:15 - Orchestrator notified
15:33:42 - builder-2 contacted
15:35:18 - builder-2 acknowledged
15:54:32 - Partial interface delivered
15:55:05 - Blocker resolved

Resolution time: 25 minutes
Impact: Minimal (partial unblocking achieved)
Status: ✅ RESOLVED

Rationale: Demonstrates blocker escalation, orchestrator mediation, and partial unblocking strategy

Quality Standards

Output Requirements

All messages must conform to defined JSON schema with required fields
Message delivery must be confirmed with status tracking
Broadcast messages must reach all registered agents
Shared context updates must be atomic (locked during update)
Handoff packages must include complete artifacts and acceptance criteria

Performance Requirements

Message delivery latency: ≤1 second for direct messages
Inbox polling frequency: Every 30 seconds during active work
Shared context lock acquisition: ≤30 seconds wait time
Broadcast delivery: ≤5 seconds to all agents
Event processing: ≤2 seconds from publish to subscriber delivery

Integration Requirements

Integrates with parallel-executor-skill for agent coordination
Uses worktree-manager-skill worktree paths for routing
Reports communication metrics to orchestrator dashboards
Supports MULTI_AGENT_PLAN.md handoff protocols

Common Pitfalls

Pitfall 1: Message Loss from Unreliable Delivery

Issue: Messages sent but never received by agent Why it happens: File system errors, permissions issues, process crashes Solution:

Implement message acknowledgment requirement
Retry unacknowledged messages after timeout
Log all sent messages to audit trail
Implement dead letter queue for failed deliveries

Pitfall 2: Race Conditions in Shared Context Updates

Issue: Concurrent updates overwriting each other Why it happens: Multiple agents updating context simultaneously without locking Solution:

Always acquire lock before updating shared context
Implement timeout for lock acquisition
Use optimistic locking with version numbers
Consider event sourcing for append-only updates

Pitfall 3: Conversation Fragmentation

Issue: Related messages not grouped, hard to follow conversation Why it happens: Missing conversation_id tracking Solution:

Always use conversation_id for related messages
Link responses to original message with in_reply_to
Implement conversation view in message reader
Provide conversation summary tools

Pitfall 4: Broadcast Spam

Issue: Too many broadcast messages, agents overwhelmed Why it happens: Every minor update broadcast to all agents Solution:

Reserve broadcasts for critical decisions only
Use event subscriptions for opt-in notifications
Implement message priority and filtering
Batch related updates into single broadcast

Pitfall 5: Stale Shared Context

Issue: Agents reading outdated shared context Why it happens: Context updated but agents don't refresh Solution:

Broadcast context_updated event when context changes
Agents should re-read context when event received
Display last_updated timestamp prominently
Implement cache invalidation in context readers

Integration with Command & Control

Related Agents

Orchestrator Agent: Subscribes to all events, mediates conflicts, facilitates communication
Worker Agents: Send/receive messages, update shared context, participate in handoffs

Related Commands

/handoff: Uses agent-communication for handoff documentation
/start-session: Initializes message bus for session
/close-session: Archives message history

Related Skills

parallel-executor-skill: Uses messaging for agent coordination (dependency)
multi-agent-planner-skill: Defines communication protocols in plan
worktree-manager-skill: Provides agent locations for routing

MCP Dependencies

Filesystem MCP: Message file creation, inbox polling
Process MCP: Background message polling processes

Orchestration Notes

Invoked by: parallel-executor-skill (enable coordination), orchestrator agent
Invokes: None (leaf skill)
Used throughout: All multi-agent workflows

Troubleshooting

Issue: Messages Not Being Delivered

Symptoms: Agent sends message but recipient never receives Diagnosis:

bash

# Check if message file created
ls -la .agent-messages/inbox/recipient-agent/

# Check message file permissions
ls -l .agent-messages/inbox/recipient-agent/*.json

# Check registry
jq '.agents["recipient-agent"]' .agent-messages/registry.json

Solution:

Verify recipient agent registered
Check file permissions (should be readable)
Verify inbox directory exists
Check message file format (must be valid JSON)

Issue: Shared Context Lock Timeout

Symptoms: Cannot acquire lock on shared context after 30 seconds Diagnosis:

bash

# Check lock status
jq . .agent-messages/shared-context/api-contracts.md.lock

# Check who holds lock
jq -r '.locked_by' .agent-messages/shared-context/api-contracts.md.lock

Solution:

bash

# If locked by crashed agent, manually release
jq '.locked = false | .locked_by = null | .locked_at = null' \
  .agent-messages/shared-context/api-contracts.md.lock > temp.json
mv temp.json .agent-messages/shared-context/api-contracts.md.lock

Issue: Event Subscriptions Not Working

Symptoms: Agent subscribed to event but not receiving notifications Diagnosis:

bash

# Check subscriptions
jq '.event_subscriptions' .agent-messages/registry.json

# Check event directory
ls -la .agent-messages/events/TaskCompleted/

Solution:

Verify agent subscribed to correct event type
Check event publishing actually creates files
Verify event delivery copies to agent inbox
Re-subscribe agent to event

Version History

1.0.0 (2025-11-29): Initial release
- Inter-agent messaging with multiple protocols
- Shared context management with locking
- Event broadcasting and subscriptions
- Handoff documentation automation
- Message persistence and archival
- Conversation threading and tracking
- Integration with orchestration ecosystem

Search AI Tools

Install this agent skill to your Project

SKILL.md

Agent Communication Skill

Description

When to Use This Skill

When NOT to Use This Skill

Prerequisites

Workflow

Phase 1: Communication Infrastructure Setup

Step 1.1: Initialize Message Bus

Step 1.2: Register Agents

Step 1.3: Define Message Schema

Phase 2: Message Sending and Routing

Step 2.1: Send Direct Message

Step 2.2: Broadcast Message

Step 2.3: Publish Event

Phase 3: Message Receiving and Processing

Step 3.1: Check Inbox for New Messages

Step 3.2: Read and Process Message

Step 3.3: Respond to Message

Phase 4: Shared Context Management

Step 4.1: Create Shared Context Document

Step 4.2: Update Shared Context

Phase 5: Handoff Documentation

Step 5.1: Create Handoff Package

Step 5.2: Accept Handoff

Examples

Example 1: Interface Contract Negotiation

Example 2: Architectural Decision Broadcast

Example 3: Blocker Escalation and Resolution

Quality Standards

Output Requirements

Performance Requirements

Integration Requirements

Common Pitfalls

Pitfall 1: Message Loss from Unreliable Delivery

Pitfall 2: Race Conditions in Shared Context Updates

Pitfall 3: Conversation Fragmentation

Pitfall 4: Broadcast Spam

Pitfall 5: Stale Shared Context

Integration with Command & Control

Related Agents

Related Commands

Related Skills

MCP Dependencies

Orchestration Notes

Troubleshooting

Issue: Messages Not Being Delivered

Issue: Shared Context Lock Timeout

Issue: Event Subscriptions Not Working

Version History