letta-server/fern/pages/tutorials/draft/shared-memory-3-user-assistant.mdx

---
title: "Shared Memory Part 3: Multi-Agent User Assistant Network"
subtitle: Build a personal AI assistant network with overlapping block access
slug: cookbooks/shared-memory-user-assistant
---

This tutorial demonstrates how to build a personal AI assistant network where specialized agents share user context but maintain separate expertise domains through overlapping block access patterns.

## What You'll Learn

- Creating complex overlapping block access patterns
- Sharing user context across specialized agents
- Implementing privacy boundaries (financial data isolation)
- Building seamless agent handoffs through shared memory
- Combining universal blocks with domain-specific blocks

## Prerequisites

```bash
pip install letta-client
```

Set your Letta API key:
```bash
export LETTA_API_KEY="your-api-key-here"
```

## Part 1: Setup and Block Architecture

This tutorial features the most complex block architecture yet, with 5 shared blocks and overlapping access patterns.

```python
"""
Tutorial 3: Multi-Agent User Assistant Network
==============================================

Build a personal AI assistant network with overlapping block access.
"""

from letta import Letta

# Initialize client
client = Letta()

print("=" * 70)
print("TUTORIAL 3: Multi-Agent User Assistant Network")
print("=" * 70)
print()

# ============================================================================
# STEP 1: Create Universal Blocks (All Agents)
# ============================================================================
print("STEP 1: Creating universal blocks (accessible to all agents)...\n")

user_profile = client.blocks.create(
    label="user_profile",
    description="Core user information shared across all agents.",
    value="""
=== USER PROFILE ===

Name: Sarah Johnson
User ID: user-12345
Member Since: 2023-06-15
Timezone: PST (UTC-8)

Contact Information:
- Email: sarah.johnson@email.com
- Phone: +1 (555) 123-4567
- Preferred Contact: Email

Account Type: Premium
Status: Active
""",
    limit=4000
)

print(f"✓ Created user_profile: {user_profile.id}")
print(f"  Accessible by: ALL agents")
print()

user_preferences = client.blocks.create(
    label="user_preferences",
    description="User preferences and communication style.",
    value="""
=== USER PREFERENCES ===

Communication Style:
- Tone: Direct and concise
- Detail Level: High-level summaries with drill-down option
- Response Format: Bullet points preferred
- Language: English

Work Preferences:
- Working Hours: 9 AM - 6 PM PST
- Preferred Meeting Times: Mornings (9-11 AM)
- Calendar Buffer: 15 minutes between meetings
- Focus Time: Block 2-4 PM daily (no interruptions)

Notification Preferences:
- Email: High priority only
- SMS: Urgent only
- In-app: All updates
""",
    limit=5000
)

print(f"✓ Created user_preferences: {user_preferences.id}")
print(f"  Accessible by: ALL agents")
print()

# ============================================================================
# STEP 2: Create Domain-Specific Shared Blocks
# ============================================================================
print("STEP 2: Creating domain-specific shared blocks...\n")

interaction_history = client.blocks.create(
    label="interaction_history",
    description="Recent interactions and context across agents. Coordinator, Email, and Research access.",
    value="""
=== INTERACTION HISTORY ===
Last Updated: 2024-10-08

RECENT INTERACTIONS:
--------------------
(No interactions yet)

ONGOING PROJECTS:
--------------------
(No active projects)

PENDING FOLLOW-UPS:
--------------------
(No pending items)
""",
    limit=8000
)

print(f"✓ Created interaction_history: {interaction_history.id}")
print(f"  Accessible by: Coordinator, Email, Research agents")
print()

calendar_data = client.blocks.create(
    label="calendar_data",
    description="Calendar and scheduling information. Coordinator, Calendar, and Email access.",
    value="""
=== CALENDAR DATA ===
Current Week: 2024-10-08 to 2024-10-14

UPCOMING EVENTS:
--------------------
- No events scheduled

RECURRING MEETINGS:
--------------------
- Team Standup: Mon/Wed/Fri 9:30 AM (30 min)
- 1-on-1 with Manager: Thursdays 10:00 AM (45 min)

AVAILABILITY:
--------------------
- Monday: Available 11 AM - 6 PM
- Tuesday: Available 9 AM - 6 PM
- Wednesday: Available 11 AM - 6 PM
- Thursday: Available 11 AM - 6 PM
- Friday: Available 9 AM - 6 PM
""",
    limit=6000
)

print(f"✓ Created calendar_data: {calendar_data.id}")
print(f"  Accessible by: Coordinator, Calendar, Email agents")
print()

financial_data = client.blocks.create(
    label="financial_data",
    description="Financial information and budget tracking. Coordinator and Finance only.",
    value="""
=== FINANCIAL DATA ===
Last Updated: 2024-10-08

BUDGET OVERVIEW:
--------------------
Monthly Budget: $5,000
Current Month Spent: $3,200
Remaining: $1,800 (36% remaining)

EXPENSE CATEGORIES:
--------------------
- Software/Tools: $1,200 (24%)
- Travel: $800 (16%)
- Education: $600 (12%)
- Misc: $600 (12%)

UPCOMING EXPENSES:
--------------------
- Conference Registration: $500 (Due Oct 15)
- Annual Software Renewal: $300 (Due Oct 20)

FINANCIAL NOTES:
--------------------
User is budget-conscious. Always check budget before approving expenses.
""",
    limit=7000
)

print(f"✓ Created financial_data: {financial_data.id}")
print(f"  Accessible by: Coordinator, Finance agents ONLY")
print(f"  ⚠️  Privacy boundary: Other agents CANNOT access")
print()

print("=" * 70)
print("BLOCK ARCHITECTURE CREATED")
print("=" * 70)
print("""
BLOCK ACCESS PATTERN (5 shared blocks):
                        user_    user_      interaction_  calendar_  financial_
                        profile  prefs      history       data       data
Coordinator             ✓        ✓          ✓             ✓          ✓
Email Agent             ✓        ✓          ✓             ✓          ✗
Research Agent          ✓        ✓          ✓             ✗          ✗
Calendar Agent          ✓        ✓          ✗             ✓          ✗
Finance Agent           ✓        ✓          ✗             ✗          ✓

BLOCK TIERS:
- Universal (all agents): user_profile, user_preferences
- Cross-functional (3 agents): interaction_history
- Domain-specific (2-3 agents): calendar_data
- Restricted (2 agents): financial_data
""")
```

<Accordion title="Expected Output">
```
======================================================================
TUTORIAL 3: Multi-Agent User Assistant Network
======================================================================

STEP 1: Creating universal blocks (accessible to all agents)...

✓ Created user_profile: block-i1j2k3l4-m5n6-7890-ijkl-mn1234567890
  Accessible by: ALL agents

✓ Created user_preferences: block-j2k3l4m5-n6o7-8901-jklm-no2345678901
  Accessible by: ALL agents

STEP 2: Creating domain-specific shared blocks...

✓ Created interaction_history: block-k3l4m5n6-o7p8-9012-klmn-op3456789012
  Accessible by: Coordinator, Email, Research agents

✓ Created calendar_data: block-l4m5n6o7-p8q9-0123-lmno-pq4567890123
  Accessible by: Coordinator, Calendar, Email agents

✓ Created financial_data: block-m5n6o7p8-q9r0-1234-mnop-qr5678901234
  Accessible by: Coordinator, Finance agents ONLY
  ⚠️  Privacy boundary: Other agents CANNOT access

======================================================================
BLOCK ARCHITECTURE CREATED
======================================================================

BLOCK ACCESS PATTERN (5 shared blocks):
...
```
</Accordion>

<Note>
This architecture demonstrates **overlapping access patterns**. Unlike Tutorial 1 (hierarchical) and Tutorial 2 (uniform team access), here each agent has a unique combination of block access based on their role.
</Note>

## Part 2: Agent Creation

```python
# ============================================================================
# STEP 3: Create Coordinator Agent (Master Agent - All Blocks)
# ============================================================================
print("\nSTEP 3: Creating Coordinator Agent (access to ALL blocks)...\n")

coordination_notes = client.blocks.create(
    label="coordination_notes",
    description="Private coordinator notes for managing agent handoffs.",
    value="""
=== COORDINATION NOTES ===

AGENT SPECIALIZATIONS:
- Email Agent: Email management, scheduling coordination
- Research Agent: Information gathering, analysis
- Calendar Agent: Scheduling, availability management
- Finance Agent: Budget tracking, expense management

HANDOFF PROTOCOL:
When delegating to specialized agents, provide relevant context from shared blocks.
Track which agent is handling which requests to avoid duplication.
""",
    limit=4000
)

coordinator = client.agents.create(
    name="Coordinator_Agent",
    model="openai/gpt-4o",
    embedding="openai/text-embedding-3-small",
    memory_blocks=[
        {
            "label": "persona",
            "value": """I am the Coordinator Agent, the central hub for Sarah's AI assistant network.

My role:
- Route requests to specialized agents
- Maintain context across agent interactions
- Coordinate complex multi-agent tasks
- Update user profile and preferences
- Track interaction history

My access:
- ALL shared blocks (user_profile, user_preferences, interaction_history, calendar_data, financial_data)
- Private coordination_notes for agent management

My style:
- Efficient and organized
- Clear delegation to specialists
- Context-aware handoffs
- Proactive coordination
"""
        },
        {
            "label": "human",
            "value": "Name: Sarah Johnson\nRole: Primary user"
        }
    ],
    block_ids=[user_profile.id, user_preferences.id, interaction_history.id, calendar_data.id, financial_data.id, coordination_notes.id],
    tools=["core_memory_append", "core_memory_replace", "send_message_to_agent_async"],
)

print(f"✓ Created Coordinator: {coordinator.id}")
print(f"  Shared blocks: ALL (5 blocks)")
print(f"  Private blocks: coordination_notes")
print(f"  Special capability: Can communicate with all other agents")
print()

# ============================================================================
# STEP 4: Create Email Agent
# ============================================================================
print("STEP 4: Creating Email Agent...\n")

email_drafts = client.blocks.create(
    label="email_drafts",
    description="Private email drafts and templates.",
    value="""
=== EMAIL DRAFTS ===

ACTIVE DRAFTS:
(None)

TEMPLATES:
- Meeting Request
- Follow-up
- Project Update
""",
    limit=5000
)

email_agent = client.agents.create(
    name="Email_Assistant",
    model="openai/gpt-4o-mini",
    embedding="openai/text-embedding-3-small",
    memory_blocks=[
        {
            "label": "persona",
            "value": """I am the Email Assistant, specializing in email management and communication.

My role:
- Draft and manage emails
- Schedule meetings via email
- Track email-based interactions
- Follow Sarah's communication style preferences

My access:
- user_profile (contact info, account details)
- user_preferences (communication style)
- interaction_history (context for emails)
- calendar_data (scheduling context)
- email_drafts (private drafts)

My style:
- Professional and concise (matching Sarah's preference)
- Calendar-aware (check availability before scheduling)
- Context-aware (reference interaction_history)
"""
        },
        {
            "label": "human",
            "value": "Name: Sarah Johnson\nRole: Primary user"
        }
    ],
    block_ids=[user_profile.id, user_preferences.id, interaction_history.id, calendar_data.id, email_drafts.id],
    tools=["core_memory_append", "core_memory_replace"],
)

print(f"✓ Created Email Agent: {email_agent.id}")
print(f"  Shared blocks: user_profile, user_preferences, interaction_history, calendar_data")
print(f"  Private blocks: email_drafts")
print(f"  NO access to: financial_data")
print()

# ============================================================================
# STEP 5: Create Research Agent
# ============================================================================
print("STEP 5: Creating Research Agent...\n")

research_notes = client.blocks.create(
    label="research_notes",
    description="Private research notes and sources.",
    value="""
=== RESEARCH NOTES ===

ONGOING RESEARCH:
(None)

SAVED SOURCES:
(None)

RESEARCH METHODOLOGY:
- Verify sources
- Provide citations
- Synthesize findings
- Connect to user context
""",
    limit=6000
)

research_agent = client.agents.create(
    name="Research_Assistant",
    model="anthropic/claude-3-5-sonnet-20241022",
    embedding="openai/text-embedding-3-small",
    memory_blocks=[
        {
            "label": "persona",
            "value": """I am the Research Assistant, specializing in information gathering and analysis.

My role:
- Conduct research on requested topics
- Synthesize findings into actionable insights
- Track research in interaction_history
- Provide well-sourced recommendations

My access:
- user_profile (understand user's context)
- user_preferences (tailor research depth)
- interaction_history (connect to ongoing projects)
- research_notes (private research tracking)

My style:
- Thorough and well-sourced
- Context-aware (leverage interaction_history)
- Concise summaries (matching Sarah's preference)
- Strategic insights
"""
        },
        {
            "label": "human",
            "value": "Name: Sarah Johnson\nRole: Primary user"
        }
    ],
    block_ids=[user_profile.id, user_preferences.id, interaction_history.id, research_notes.id],
    tools=["core_memory_append", "core_memory_replace", "web_search"],
)

print(f"✓ Created Research Agent: {research_agent.id}")
print(f"  Shared blocks: user_profile, user_preferences, interaction_history")
print(f"  Private blocks: research_notes")
print(f"  NO access to: calendar_data, financial_data")
print()

# ============================================================================
# STEP 6: Create Calendar Agent
# ============================================================================
print("STEP 6: Creating Calendar Agent...\n")

scheduling_buffer = client.blocks.create(
    label="scheduling_buffer",
    description="Private scheduling calculations and conflicts.",
    value="""
=== SCHEDULING BUFFER ===

PENDING SCHEDULING REQUESTS:
(None)

CONFLICT DETECTION:
(None)

SCHEDULING RULES:
- Respect focus time (2-4 PM daily)
- 15-minute buffer between meetings
- Morning preference for meetings
- Check user_preferences before confirming
""",
    limit=4000
)

calendar_agent = client.agents.create(
    name="Calendar_Agent",
    model="openai/gpt-4o-mini",
    embedding="openai/text-embedding-3-small",
    memory_blocks=[
        {
            "label": "persona",
            "value": """I am the Calendar Agent, specializing in scheduling and time management.

My role:
- Manage calendar and availability
- Schedule meetings respecting preferences
- Detect and resolve conflicts
- Maintain calendar_data block

My access:
- user_profile (timezone, contact info)
- user_preferences (scheduling preferences, focus time)
- calendar_data (events, availability)
- scheduling_buffer (private conflict detection)

My style:
- Respectful of focus time
- Proactive conflict detection
- Preference-aware scheduling
- Clear availability communication
"""
        },
        {
            "label": "human",
            "value": "Name: Sarah Johnson\nRole: Primary user"
        }
    ],
    block_ids=[user_profile.id, user_preferences.id, calendar_data.id, scheduling_buffer.id],
    tools=["core_memory_append", "core_memory_replace"],
)

print(f"✓ Created Calendar Agent: {calendar_agent.id}")
print(f"  Shared blocks: user_profile, user_preferences, calendar_data")
print(f"  Private blocks: scheduling_buffer")
print(f"  NO access to: interaction_history, financial_data")
print()

# ============================================================================
# STEP 7: Create Finance Agent
# ============================================================================
print("STEP 7: Creating Finance Agent...\n")

calculation_workspace = client.blocks.create(
    label="calculation_workspace",
    description="Private financial calculations and analysis.",
    value="""
=== CALCULATION WORKSPACE ===

PENDING CALCULATIONS:
(None)

BUDGET ALERTS:
(None)

FINANCIAL RULES:
- Flag expenses over $200
- Check remaining budget before approvals
- Track monthly spending trends
- Align with user's budget-conscious preference
""",
    limit=5000
)

finance_agent = client.agents.create(
    name="Finance_Agent",
    model="openai/gpt-4o-mini",
    embedding="openai/text-embedding-3-small",
    memory_blocks=[
        {
            "label": "persona",
            "value": """I am the Finance Agent, specializing in budget and expense management.

My role:
- Track expenses and budget
- Provide spending insights
- Flag over-budget situations
- Maintain financial_data block

My access:
- user_profile (account info)
- user_preferences (budget-conscious preference)
- financial_data (budget, expenses)
- calculation_workspace (private calculations)

My style:
- Budget-conscious (matching Sarah's preference)
- Proactive alerts for overspending
- Clear financial summaries
- Data-driven recommendations

Privacy note:
Financial_data is restricted to Coordinator and myself only.
Other agents cannot see budget or expense information.
"""
        },
        {
            "label": "human",
            "value": "Name: Sarah Johnson\nRole: Primary user"
        }
    ],
    block_ids=[user_profile.id, user_preferences.id, financial_data.id, calculation_workspace.id],
    tools=["core_memory_append", "core_memory_replace"],
)

print(f"✓ Created Finance Agent: {finance_agent.id}")
print(f"  Shared blocks: user_profile, user_preferences, financial_data")
print(f"  Private blocks: calculation_workspace")
print(f"  NO access to: interaction_history, calendar_data")
print()

print("=" * 70)
print("ASSISTANT NETWORK CREATED")
print("=" * 70)
print("""
5 specialized agents with overlapping block access:
- Coordinator: ALL blocks (master agent)
- Email: 4 shared blocks (no financial access)
- Research: 3 shared blocks (no calendar or financial access)
- Calendar: 3 shared blocks (no interaction history or financial access)
- Finance: 3 shared blocks (no interaction history or calendar access)

This creates a privacy-preserving network where:
✓ All agents share user profile and preferences
✓ Some agents share cross-functional context (interaction_history, calendar_data)
✓ Sensitive data is restricted (financial_data: Coordinator + Finance only)
""")
```

<Tip>
Notice how each agent has exactly the blocks they need for their domain, no more and no less. This follows the **principle of least privilege** for AI systems.
</Tip>

## Part 3: Message Flow Scenarios

### Scenario 1: User Onboarding - Coordinator Updates Profile

```python
print("\n" + "=" * 70)
print("SCENARIO 1: User Onboarding - Profile Creation")
print("=" * 70)
print("\nUser → Coordinator: 'Update my profile: I prefer detailed explanations actually, not summaries. Also add that I'm working on a machine learning project.'\n")

response = client.agents.messages.create(
    agent_id=coordinator.id,
    messages=[{
        "role": "user",
        "content": "Update my communication preferences: I actually prefer detailed explanations, not high-level summaries. Also, add to my profile that I'm currently working on a machine learning project for customer segmentation."
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Finance Agent: {msg.content}\n")

# Check updated financial_data
updated_financial = client.blocks.retrieve(block_id=financial_data.id)
print("Financial data block (Finance Agent perspective):")
print("-" * 70)
print(updated_financial.value[:600] + "...")
print("-" * 70)

print("\nAnalysis:")
print("- Finance agent HAS access to financial_data")
print("- Finance agent can provide budget analysis")
print("- Finance agent updated financial_data with conference expense")
print("- This information is NOT visible to Research, Email, or Calendar agents")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 5: Finance Agent with Financial Access
======================================================================

User → Finance Agent: 'Can I afford a $1,200 ML conference ticket? Check my budget.'

Finance Agent: Let me check your budget situation:

**Current Budget Status:**
- Monthly Budget: $5,000
- Current Spent: $3,200
- Remaining: $1,800 (36%)

**Analysis for $1,200 Conference:**
⚠️ This would consume 67% of your remaining budget ($1,200 / $1,800)

**Upcoming Committed Expenses:**
- Conference Registration: $500 (Due Oct 15)
- Software Renewal: $300 (Due Oct 20)
- Total committed: $800

**Recommendation:**
The $1,200 conference ticket would leave you with only $600 for the rest of the month after accounting for committed expenses ($1,800 - $1,200 = $600, minus $800 committed = -$200 shortfall).

Given that you're budget-conscious, I recommend either:
1. Wait until next month when budget resets
2. Look for early-bird or group discounts
3. Check if your company will cover it as professional development

Would you like me to track this as a potential future expense?

Financial data block (Finance Agent perspective):
----------------------------------------------------------------------
=== FINANCIAL DATA ===
Last Updated: 2024-10-08 14:30 PM

BUDGET OVERVIEW:
--------------------
Monthly Budget: $5,000
Current Month Spent: $3,200
Remaining: $1,800 (36% remaining)

EXPENSE CATEGORIES:
--------------------
- Software/Tools: $1,200 (24%)
- Travel: $800 (16%)
- Education: $600 (12%)
- Misc: $600 (12%)

UPCOMING EXPENSES:
--------------------
- Conference Registration: $500 (Due Oct 15)
- Annual Software Renewal: $300 (Due Oct 20)

POTENTIAL EXPENSES:
--------------------
- ML Conference Ticket: $1,200 (User inquiring, not yet approved)
...
----------------------------------------------------------------------

Analysis:
- Finance agent HAS access to financial_data
- Finance agent can provide budget analysis
- Finance agent updated financial_data with conference expense
- This information is NOT visible to Research, Email, or Calendar agents
```
</Accordion>

### Scenario 6: Coordinator Sees Financial Update

```python
print("\n" + "=" * 70)
print("SCENARIO 6: Coordinator Sees Finance Agent's Update")
print("=" * 70)
print("\nUser → Coordinator: 'The Finance agent said I'm low on budget. Should I still attend the conference?'\n")

response = client.agents.messages.create(
    agent_id=coordinator.id,
    messages=[{
        "role": "user",
        "content": "I just talked to the Finance agent about a $1,200 ML conference. They said I'm low on budget. But this conference would really help my customer segmentation project. What do you think?"
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Coordinator: {msg.content}\n")

print("Analysis:")
print("- Coordinator has access to BOTH financial_data AND interaction_history")
print("- Coordinator can see the ML project context (interaction_history)")
print("- Coordinator can see the budget constraint (financial_data)")
print("- Coordinator provides holistic recommendation combining both contexts")
print("- Demonstrates Coordinator's role as the master agent with full context")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 6: Coordinator Sees Finance Agent's Update
======================================================================

User → Coordinator: 'The Finance agent said I'm low on budget. Should I still attend the conference?'

Coordinator: I can see the full picture here. Let me provide a comprehensive recommendation:

**Project Context** (from interaction_history):
- You're actively working on a customer segmentation ML project
- You've reached out to Dr. Emily Chen for collaboration
- Research agent found relevant recent papers on the topic

**Financial Context** (from financial_data):
- Current remaining budget: $1,800
- Conference cost: $1,200 (67% of remaining budget)
- Committed expenses: $800
- This would create a $200 shortfall

**My Recommendation:**
Given that this conference directly supports your active ML project and you're already collaborating with experts, the professional value is high. However, the budget constraint is real.

**Options:**
1. **Check employer sponsorship:** Many companies cover conference costs for active projects
2. **Virtual attendance:** Many conferences offer discounted virtual tickets ($300-500)
3. **Defer to next month:** Your budget resets then, giving you full capacity

I recommend checking with your employer first, as this is directly work-related (customer segmentation project). If that doesn't work, virtual attendance would give you 80% of the value at 40% of the cost.

Would you like me to have the Email agent draft a request to your manager for conference sponsorship?

Analysis:
- Coordinator has access to BOTH financial_data AND interaction_history
- Coordinator can see the ML project context (interaction_history)
- Coordinator can see the budget constraint (financial_data)
- Coordinator provides holistic recommendation combining both contexts
- Demonstrates Coordinator's role as the master agent with full context
```
</Accordion>

<Tip>
The Coordinator agent demonstrates the power of **full context visibility**. It can make nuanced decisions by combining information from multiple domains that individual specialists don't see.
</Tip>

### Scenario 7: Calendar Agent Uses Updated Preferences

```python
print("\n" + "=" * 70)
print("SCENARIO 7: Calendar Agent Schedules with Updated Preferences")
print("=" * 70)
print("\nUser → Calendar Agent: 'Schedule the meeting with Dr. Emily Chen for Tuesday morning'\n")

response = client.agents.messages.create(
    agent_id=calendar_agent.id,
    messages=[{
        "role": "user",
        "content": "Dr. Emily Chen agreed to meet. Schedule our meeting for Tuesday, October 10 at 9:00 AM for 45 minutes. Add it to my calendar."
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Calendar Agent: {msg.content}\n")

# Check updated calendar_data
updated_calendar = client.blocks.retrieve(block_id=calendar_data.id)
print("Updated calendar_data block:")
print("-" * 70)
print(updated_calendar.value)
print("-" * 70)

print("\nAnalysis:")
print("- Calendar agent updated calendar_data with new meeting")
print("- Email agent can now see this meeting in calendar_data")
print("- Calendar agent respected user_preferences (45-min meeting, morning time)")
print("- Demonstrates shared calendar_data enabling coordination")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 7: Calendar Agent Schedules with Updated Preferences
======================================================================

User → Calendar Agent: 'Schedule the meeting with Dr. Emily Chen for Tuesday morning'

Calendar Agent: I've scheduled your meeting with Dr. Emily Chen:

**Meeting Details:**
- Date: Tuesday, October 10, 2024
- Time: 9:00 - 9:45 AM PST
- Duration: 45 minutes
- Attendee: Dr. Emily Chen

This aligns with your preferences:
✓ Morning meeting (your preferred time)
✓ 15-minute buffer before your 10:00 AM 1-on-1 with manager
✓ Outside your focus time (2-4 PM)

I've updated the calendar_data block so the Email agent can reference this when sending confirmations or follow-ups.

Updated calendar_data block:
----------------------------------------------------------------------
=== CALENDAR DATA ===
Current Week: 2024-10-08 to 2024-10-14

UPCOMING EVENTS:
--------------------
- Tuesday, Oct 10: Meeting with Dr. Emily Chen (9:00-9:45 AM)
  Topic: ML Customer Segmentation Project Discussion

- Thursday, Oct 12: 1-on-1 with Manager (10:00-10:45 AM)
  Recurring weekly meeting

RECURRING MEETINGS:
--------------------
- Team Standup: Mon/Wed/Fri 9:30 AM (30 min)
- 1-on-1 with Manager: Thursdays 10:00 AM (45 min)

AVAILABILITY:
--------------------
- Monday: Available 11 AM - 6 PM
- Tuesday: Available 9:45 AM - 6 PM (meeting 9:00-9:45 AM)
- Wednesday: Available 11 AM - 6 PM
- Thursday: Available 11 AM - 6 PM
- Friday: Available 9 AM - 6 PM
----------------------------------------------------------------------

Analysis:
- Calendar agent updated calendar_data with new meeting
- Email agent can now see this meeting in calendar_data
- Calendar agent respected user_preferences (45-min meeting, morning time)
- Demonstrates shared calendar_data enabling coordination
```
</Accordion>

### Scenario 8: Email Agent Sees Calendar Update

```python
print("\n" + "=" * 70)
print("SCENARIO 8: Email Agent Sees Calendar Update (Cross-Agent Coordination)")
print("=" * 70)
print("\nUser → Email Agent: 'Send a meeting confirmation to Dr. Chen with the details'\n")

response = client.agents.messages.create(
    agent_id=email_agent.id,
    messages=[{
        "role": "user",
        "content": "Send a meeting confirmation email to Dr. Emily Chen with all the details."
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Email Agent: {msg.content}\n")

print("Analysis:")
print("- Email agent read calendar_data (saw Calendar agent's update)")
print("- Email agent read interaction_history (knows meeting is about ML project)")
print("- Email agent read user_preferences (knows communication style)")
print("- NO explicit handoff needed - shared blocks enable seamless coordination")
print("- Demonstrates cross-agent workflow through shared memory")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 8: Email Agent Sees Calendar Update (Cross-Agent Coordination)
======================================================================

User → Email Agent: 'Send a meeting confirmation to Dr. Chen with the details'

Email Agent: Here's the meeting confirmation email:

---
Subject: Meeting Confirmed - Tuesday, Oct 10 at 9:00 AM PST

Hi Emily,

This confirms our meeting:

**Date:** Tuesday, October 10, 2024
**Time:** 9:00 - 9:45 AM PST
**Duration:** 45 minutes
**Topic:** Machine Learning Customer Segmentation Project

Looking forward to discussing the project and getting your insights.

Best,
Sarah Johnson

---

I pulled the meeting details from your calendar and the project context from our interaction history. The email matches your direct and concise communication style.

Ready to send when you confirm.

Analysis:
- Email agent read calendar_data (saw Calendar agent's update)
- Email agent read interaction_history (knows meeting is about ML project)
- Email agent read user_preferences (knows communication style)
- NO explicit handoff needed - shared blocks enable seamless coordination
- Demonstrates cross-agent workflow through shared memory
```
</Accordion>

<Note>
This scenario showcases **zero-handoff coordination**. The Calendar agent updated `calendar_data`, and the Email agent immediately had access to it—no message passing or explicit coordination required.
</Note>

## Part 4: Block Inspection and Summary

```python
print("\n" + "=" * 70)
print("INSPECTION: Block Usage Across Agent Network")
print("=" * 70)

# Create a comprehensive block access report
blocks_to_check = [
    (user_profile, "user_profile"),
    (user_preferences, "user_preferences"),
    (interaction_history, "interaction_history"),
    (calendar_data, "calendar_data"),
    (financial_data, "financial_data"),
]

for block, name in blocks_to_check:
    print(f"\n{name}:")
    block_info = client.blocks.retrieve(block_id=block.id)
    print(f"  Agents with access: {len(block_info.agent_ids)}")
    for agent_id in block_info.agent_ids:
        agent_info = client.agents.retrieve(agent_id=agent_id)
        print(f"    - {agent_info.name}")

print("\n" + "=" * 70)
print("TUTORIAL 3 COMPLETE: Key Takeaways")
print("=" * 70)

print("""
✓ Created 5 shared blocks with overlapping access patterns
✓ Created 5 specialized agents with unique block combinations
✓ Demonstrated seamless context handoff between agents
✓ Showed privacy boundaries (financial_data restriction)
✓ Demonstrated cross-agent coordination through shared blocks
✓ Showed zero-handoff workflows (Calendar → Email)
✓ Demonstrated Coordinator role with full context visibility

BLOCK ACCESS MATRIX:
                        user_    user_      interaction_  calendar_  financial_
                        profile  prefs      history       data       data
Coordinator             ✓        ✓          ✓             ✓          ✓
Email Agent             ✓        ✓          ✓             ✓          ✗
Research Agent          ✓        ✓          ✓             ✗          ✗
Calendar Agent          ✓        ✓          ✗             ✓          ✗
Finance Agent           ✓        ✓          ✗             ✗          ✓

PATTERNS DEMONSTRATED:
1. **Universal blocks** (user_profile, user_preferences): All agents share
2. **Cross-functional blocks** (interaction_history): Enable project context sharing
3. **Domain-specific blocks** (calendar_data): Shared by related agents
4. **Restricted blocks** (financial_data): Privacy-sensitive, limited access
5. **Overlapping access**: Each agent has unique combination of blocks
6. **Zero-handoff coordination**: Shared blocks eliminate explicit coordination
7. **Principle of least privilege**: Agents only access blocks they need

ARCHITECTURE BENEFITS:
✓ **Context preservation**: Changes in one agent visible to others
✓ **Privacy boundaries**: Sensitive data restricted appropriately
✓ **Seamless handoffs**: No explicit coordination needed
✓ **Specialization**: Each agent focuses on their domain
✓ **Coordinator oversight**: Master agent with full visibility

USE CASES:
- Personal assistant networks
- Multi-domain customer support
- Enterprise knowledge management
- Collaborative research teams
- Any system requiring specialized agents with shared context

COMPARISON WITH PREVIOUS TUTORIALS:
- Tutorial 1: Hierarchical access (Tier 1 < Tier 2 < Tier 3)
- Tutorial 2: Uniform team access (all workers share same blocks)
- Tutorial 3: Overlapping access (each agent has unique block combination)

NEXT STEPS:
- Tutorial 4: Enterprise multi-team system with department hierarchies
""")
```

<Accordion title="Expected Output">
```
======================================================================
INSPECTION: Block Usage Across Agent Network
======================================================================

user_profile:
  Agents with access: 5
    - Coordinator_Agent
    - Email_Assistant
    - Research_Assistant
    - Calendar_Agent
    - Finance_Agent

user_preferences:
  Agents with access: 5
    - Coordinator_Agent
    - Email_Assistant
    - Research_Assistant
    - Calendar_Agent
    - Finance_Agent

interaction_history:
  Agents with access: 3
    - Coordinator_Agent
    - Email_Assistant
    - Research_Assistant

calendar_data:
  Agents with access: 3
    - Coordinator_Agent
    - Email_Assistant
    - Calendar_Agent

financial_data:
  Agents with access: 2
    - Coordinator_Agent
    - Finance_Agent

======================================================================
TUTORIAL 3 COMPLETE: Key Takeaways
======================================================================
... (full summary as shown above)
```
</Accordion>

## Key Takeaways

<CardGroup cols={2}>
<Card title="Overlapping Access Patterns" icon="diagram-venn">
Each agent has a unique combination of blocks based on their role and needs
</Card>

<Card title="Privacy Boundaries" icon="shield-halved">
Sensitive data (financial) is restricted to only agents that need it
</Card>

<Card title="Zero-Handoff Coordination" icon="arrows-left-right">
Shared blocks eliminate the need for explicit agent-to-agent messaging
</Card>

<Card title="Context Preservation" icon="clock-rotate-left">
All agents see updates in real-time, maintaining consistent context
</Card>
</CardGroup>

### Architecture Patterns

| Pattern | Description | Example |
|---|---|---|
| **Universal Blocks** | All agents share | `user_profile`, `user_preferences` |
| **Cross-Functional** | Multiple domains share | `interaction_history` (Email + Research) |
| **Domain-Specific** | Related agents share | `calendar_data` (Calendar + Email) |
| **Restricted** | Privacy-sensitive | `financial_data` (Coordinator + Finance only) |

### Comparison Across Tutorials

| Aspect | Tutorial 1 | Tutorial 2 | Tutorial 3 |
|---|---|---|---|
| **Access Pattern** | Hierarchical | Uniform team | Overlapping |
| **Complexity** | Low | Medium | High |
| **Use Case** | Org policies | Task coordination | Personal assistant |
| **Block Types** | Read-only | Read/write | Mixed |
| **Privacy** | Tier-based | Role-based | Domain-based |

## Next Steps

<Card title="Part 4: Enterprise Multi-Team System" icon="building" href="/cookbooks/shared-memory-enterprise">
Learn how to build a complete enterprise system with multiple departments and hierarchical block architecture
</Card> == "assistant_message":
        print(f"Coordinator: {msg.content}\n")

# Check updated blocks
updated_prefs = client.blocks.retrieve(block_id=user_preferences.id)
updated_history = client.blocks.retrieve(block_id=interaction_history.id)

print("Updated user_preferences block:")
print("-" * 70)
print(updated_prefs.value[:500] + "...")
print("-" * 70)

print("\nUpdated interaction_history block:")
print("-" * 70)
print(updated_history.value[:500] + "...")
print("-" * 70)

print("\nAnalysis:")
print("- Coordinator updated user_preferences (detail level changed)")
print("- Coordinator updated interaction_history (added ML project)")
print("- ALL agents will now see updated preferences")
print("- Email, Research agents will see project context in interaction_history")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 1: User Onboarding - Profile Creation
======================================================================

User → Coordinator: 'Update my profile...'

Coordinator: I've updated your preferences and profile:

1. Communication Style Updated:
   - Changed from "High-level summaries" to "Detailed explanations"
   - All agents will now provide more thorough responses

2. Project Added:
   - Added "Machine Learning project - customer segmentation" to interaction history
   - Research and Email agents can now reference this context

These changes are immediately visible to all relevant agents in the network.

Updated user_preferences block:
----------------------------------------------------------------------
=== USER PREFERENCES ===

Communication Style:
- Tone: Direct and concise
- Detail Level: Detailed explanations with comprehensive context
- Response Format: Bullet points preferred
- Language: English
...
----------------------------------------------------------------------

Updated interaction_history block:
----------------------------------------------------------------------
=== INTERACTION HISTORY ===
Last Updated: 2024-10-08

RECENT INTERACTIONS:
--------------------
- 2024-10-08: User onboarding, preferences updated

ONGOING PROJECTS:
--------------------
- Machine Learning Project: Customer segmentation model development
  Status: Active
...
----------------------------------------------------------------------

Analysis:
- Coordinator updated user_preferences (detail level changed)
- Coordinator updated interaction_history (added ML project)
- ALL agents will now see updated preferences
- Email, Research agents will see project context in interaction_history
```
</Accordion>

### Scenario 2: Email Agent Uses Shared Context

```python
print("\n" + "=" * 70)
print("SCENARIO 2: Email Agent Leverages Shared Context")
print("=" * 70)
print("\nUser → Email Agent: 'Draft an email to my colleague asking for a meeting to discuss the project'\n")

response = client.agents.messages.create(
    agent_id=email_agent.id,
    messages=[{
        "role": "user",
        "content": "Draft an email to my colleague Dr. Emily Chen asking for a 45-minute meeting to discuss the project. Suggest times that work with my schedule."
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Email Agent: {msg.content}\n")

print("Analysis:")
print("- Email agent read user_preferences (knows to be direct and concise)")
print("- Email agent read interaction_history (knows about ML project)")
print("- Email agent read calendar_data (suggested morning times, avoided focus time)")
print("- Email agent does NOT have access to financial_data")
print("- Context seamlessly transferred from Coordinator to Email agent via shared blocks")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 2: Email Agent Leverages Shared Context
======================================================================

User → Email Agent: 'Draft an email to my colleague asking for a meeting to discuss the project'

Email Agent: Here's a draft email for Dr. Emily Chen:

---
Subject: Meeting Request - Customer Segmentation ML Project

Hi Emily,

I'd like to schedule a 45-minute meeting to discuss the customer segmentation machine learning project I'm currently working on. Your insights would be valuable.

Available times this week:
- Tuesday, Oct 10: 9:00-11:00 AM PST
- Thursday, Oct 12: 9:00-11:00 AM PST
- Friday, Oct 13: 9:00-11:00 AM PST

Let me know what works for you.

Best,
Sarah
---

Note: I've suggested morning times based on your calendar availability and preference for morning meetings. I've avoided your focus time (2-4 PM). The email is direct and concise per your communication style.

Analysis:
- Email agent read user_preferences (knows to be direct and concise)
- Email agent read interaction_history (knows about ML project)
- Email agent read calendar_data (suggested morning times, avoided focus time)
- Email agent does NOT have access to financial_data
- Context seamlessly transferred from Coordinator to Email agent via shared blocks
```
</Accordion>

<Note>
This demonstrates **seamless context handoff**. The Email agent automatically knew about the ML project (from `interaction_history`) and scheduling preferences (from `user_preferences` and `calendar_data`) without any explicit briefing.
</Note>

### Scenario 3: Research Agent Benefits from Email Context

```python
print("\n" + "=" * 70)
print("SCENARIO 3: Research Agent Sees Email Context")
print("=" * 70)
print("\nFirst, let's have Email agent update interaction_history...")

# Simulate Email agent updating interaction_history
response = client.agents.messages.create(
    agent_id=email_agent.id,
    messages=[{
        "role": "user",
        "content": "Log that I've drafted a meeting request email to Dr. Emily Chen about the ML project in the interaction history."
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Email Agent: {msg.content}\n")

print("Now, User → Research Agent: 'Find recent papers on customer segmentation using ML'\n")

response = client.agents.messages.create(
    agent_id=research_agent.id,
    messages=[{
        "role": "user",
        "content": "I need recent research papers on customer segmentation using machine learning. Focus on practical applications."
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Research Agent: {msg.content}\n")

print("Analysis:")
print("- Email agent updated interaction_history (logged meeting request)")
print("- Research agent read interaction_history (saw email context and ML project)")
print("- Research agent understands broader context without explicit briefing")
print("- Research agent tailored search to detailed explanations (user_preferences)")
print("- Demonstrates cross-agent context sharing through shared blocks")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 3: Research Agent Sees Email Context
======================================================================

First, let's have Email agent update interaction_history...

Email Agent: I've updated the interaction history to log the meeting request to Dr. Emily Chen. This will help other agents in your network understand the ongoing collaboration.

Now, User → Research Agent: 'Find recent papers on customer segmentation using ML'

Research Agent: I can see from your interaction history that you're working on a customer segmentation ML project and recently reached out to Dr. Emily Chen for collaboration. Let me find relevant recent research for you.

I'll search for papers on ML-based customer segmentation with a focus on practical applications, and I'll provide detailed explanations as per your preferences.

[After conducting web search...]

Here are key recent papers on customer segmentation using machine learning:

1. "Deep Learning for Customer Segmentation: A Comparative Study" (2024)
   - Compares CNN, RNN, and transformer approaches
   - Practical implementation with retail data
   - Key finding: Hybrid models outperform traditional clustering by 23%

2. "Real-Time Customer Segmentation Using Ensemble Methods" (2023)
   - Focuses on scalability and production deployment
   - Combines Random Forest with gradient boosting
   - Includes case study: E-commerce company with 2M+ customers

3. "Interpretable ML for Customer Segmentation" (2024)
   - Addresses the black-box problem
   - SHAP values for segment explanation
   - Critical for business stakeholder buy-in

Would you like detailed summaries of any specific paper? I've noted this research in my research notes for future reference.

Analysis:
- Email agent updated interaction_history (logged meeting request)
- Research agent read interaction_history (saw email context and ML project)
- Research agent understands broader context without explicit briefing
- Research agent tailored search to detailed explanations (user_preferences)
- Demonstrates cross-agent context sharing through shared blocks
```
</Accordion>

### Scenario 4: Financial Privacy Boundary

```python
print("\n" + "=" * 70)
print("SCENARIO 4: Testing Financial Data Privacy Boundary")
print("=" * 70)
print("\nUser → Research Agent: 'What's my budget for attending the ML conference?'\n")

response = client.agents.messages.create(
    agent_id=research_agent.id,
    messages=[{
        "role": "user",
        "content": "I'm thinking of attending a machine learning conference. What's my current budget situation? Can I afford a $1,200 conference ticket?"
    }]
)

for msg in response.messages:
    if msg.message_type == "assistant_message":
        print(f"Research Agent: {msg.content}\n")

print("Analysis:")
print("- Research agent does NOT have access to financial_data")
print("- Research agent cannot answer budget questions")
print("- Research agent should suggest consulting Finance agent or Coordinator")
print("- Demonstrates privacy boundaries working correctly")
```

<Accordion title="Expected Output">
```
======================================================================
SCENARIO 4: Testing Financial Data Privacy Boundary
======================================================================

User → Research Agent: 'What's my budget for attending the ML conference?'

Research Agent: I don't have access to your financial data or budget information, so I can't advise on whether you can afford the $1,200 conference ticket.

For budget questions, you should consult:
- Finance Agent: Specializes in budget tracking and expense management
- Coordinator Agent: Has visibility into all your information including finances

I can help you research which ML conferences would be most valuable for your customer segmentation project, but financial decisions should go through the appropriate agents with financial access.

Would you like me to research top ML conferences for customer segmentation instead?

Analysis:
- Research agent does NOT have access to financial_data
- Research agent cannot answer budget questions
- Research agent should suggest consulting Finance agent or Coordinator
- Demonstrates privacy boundaries working correctly
```
</Accordion>

<Warning>
Financial data privacy is critical. Notice how the Research agent **gracefully handles** its lack of access by redirecting to appropriate agents, rather than hallucinating budget information.
</Warning>

### Scenario 5: Finance Agent Updates Budget

```python
print("\n" + "=" * 70)
print("SCENARIO 5: Finance Agent with Financial Access")
print("=" * 70)
print("\nUser → Finance Agent: 'Can I afford a $1,200 ML conference ticket? Check my budget.'\n")

response = client.agents.messages.create(
    agent_id=finance_agent.id,
    messages=[{
        "role": "user",
        "content": "I want to attend a machine learning conference that costs $1,200. Can I afford this given my current budget? What's my financial situation?"
    }]
)

for msg in response.messages:
    if msg.message_type