How to AWS Strands Swarm share memory for context exchange, and built-in telemetry?

both of your questions relate to the same underlying multi-agent architecture capabilities. Here are the precise technical answers to your points:

1. Combining Graph and Swarm Patterns Yes, you can use both in the same system via nested patterns. You can define an entire "Swarm" (a group of specialized agents dynamically routing tasks) as a single execution node within a broader, deterministic "Graph" workflow.

2. Graph Pattern Node Limits There is no hard-coded limit to the number of nodes in the graph pattern. The size and depth of your graph are only constrained by practical execution limits, primarily:

• The Amazon Bedrock 30-minute execution timeout limit.
• Maximum token limits of the underlying LLMs.
• Your configured max-iteration safeguards (to prevent infinite loops).

3. Shared Memory & Context Exchange

• Graph Pattern: Context is explicitly exchanged by passing a state dictionary (e.g., invocation_state) from node to node.
• Swarm Pattern: Agents share a unified conversation history. When a "handoff" occurs, the receiving agent inherits the full context of the previous steps.
• Persistence: For cross-session memory, both patterns can be integrated with Amazon Bedrock Memory, which handles short-term (session) and long-term context retention.

4. Built-in Telemetry Telemetry and observability are built into the execution results. Upon completion, the framework returns detailed traces, including the sequence of invoked agents, state transitions, individual outputs, latency, and precise token usage metrics.

Relevant Resources for Implementation:

• AWS Samples: For code-level implementations of swarms and graphs, refer to the Amazon Bedrock Samples GitHub Repository, which includes Pythonic API patterns for agent routing, tools, and workflows.