[FEATURE] Multi-Agent Session Manager for Graph and Swarm Patterns

[yonib05]

Problem Statement

Strands builders working with multi-agent patterns are requesting session management capabilities that extend beyond single-agent scenarios. Based on community feedback, developers need:

• Multi-Agent Session Persistence: Builders want to maintain conversation history and state across all agents in complex multi-agent patterns (Swarms, Graphs, Handoffs, Agents-as-Tools)
• Cross-Agent Context Sharing: The community is asking for shared session state that enables agents to access context from other agents within the same multi-agent execution
• Pattern-Aware Session Management: Developers want session managers that understand multi-agent coordination events like handoffs, swarm interactions, and graph transitions
• Unified Session Recovery: Requests for the ability to restore entire multi-agent workflows with all agent states and inter-agent relationships intact

Proposed Solution

We propose extending the existing session management system to support multi-agent pattern sessions, enabling builders to persist and restore complex multi-agent workflows with full context preservation.

Multi-Agent Session Management

from strands.multiagent import Swarm, GraphBuilder
from strands.session.file_session_manager import FileSessionManager

# Session manager automatically handles multiple agents
session_manager = FileSessionManager(session_id="research-team-session")

# All agents in swarm share session context
research_swarm = Swarm(
    agents=[researcher, analyst, writer],
    session_manager=session_manager
)

# Cross-agent context is automatically preserved
result = research_swarm.execute("Analyze market trends")
# All agent interactions, handoffs, and shared context persisted

Graph Pattern Session Persistence

from strands.session.s3_session_manager import S3SessionManager

# Session manager tracks graph execution state
session_manager = S3SessionManager(
    session_id="workflow-session-456",
    bucket="multi-agent-sessions"
)

builder = GraphBuilder()
builder.add_node(analyzer_agent, "analyze")
builder.add_node(processor_agent, "process")
builder.add_node(reviewer_agent, "review")
builder.set_session_manager(session_manager)

workflow = builder.build()

# Graph execution state and node transitions persisted
result = workflow.execute("Process customer feedback")
# Can resume from any point if interrupted

Handoff Session Continuity

from strands.multiagent import handoff_to_agent
from strands.session.file_session_manager import FileSessionManager

session_manager = FileSessionManager(session_id="support-session")

sales_agent = Agent(
    name="sales",
    session_manager=session_manager,
    tools=[handoff_to_agent]
)

technical_agent = Agent(
    name="technical", 
    session_manager=session_manager,
    tools=[handoff_to_agent]
)

# Handoff context automatically preserved in shared session
# technical_agent can access full conversation history from sales_agent

Agents-as-Tools Session Integration

# Specialist agents share session with orchestrator
research_assistant = Agent(
    name="research_assistant",
    session_manager=session_manager,
    tools=[web_search, memory]
)

orchestrator = Agent(
    name="orchestrator",
    session_manager=session_manager,
    tools=[research_assistant, product_assistant]
)

# All agent interactions stored in unified session
result = orchestrator("Plan a trip to Amsterdam")
# Specialist consultations preserved for future reference

Use Case

Workflow Resumption

• Interrupted Execution: Resume complex multi-agent workflows from the exact point of interruption
• Deployment Continuity: Maintain session state across application restarts and deployments
• Error Recovery: Restore multi-agent state after system failures or timeouts

Distributed Multi-Agent Systems

• Cloud Persistence: Store multi-agent sessions in S3 for distributed deployments
• Session Sharing: Multiple application instances access the same multi-agent session
• Scalable State Management: Handle large-scale multi-agent systems with persistent state

Long-Running Processes

• Multi-Day Workflows: Persist complex research or analysis tasks that span multiple sessions
• Iterative Refinement: Agents build upon previous work across multiple interactions
• Progressive Enhancement: Add new agents to existing workflows while preserving context

Alternatives Solutions

No response

Additional Context

• Multi-agent sessions extend the existing Session, SessionAgent, and SessionMessage data models
• New SessionPattern model will track multi-agent coordination state and relationships
• Cross-agent message access will be controlled through session-scoped permissions
• Backward compatibility maintained with existing single-agent session management