AGENTREL01-BP01 Implement a resilient messaging layer
Direct agent-to-agent calls couple failure modes. When one agent fails, everything downstream fails with it. A messaging layer with persistence, retry, and dead-letter handling absorbs transient faults and lets workflows resume from where they stopped.
Desired outcome:
-
Your agents communicate through an intermediary messaging layer with persistence, retry, and dead-letter handling rather than direct synchronous calls.
-
You have durable workflow state that survives the restart or loss of any single component.
-
You can trace every agent message across synchronous and asynchronous boundaries.
Common anti-patterns:
-
Wiring agents together through direct synchronous calls, so a single failure cascades through every dependent agent.
-
Running messaging infrastructure without persistence, making workflow recovery impossible after a component outage.
-
Treating every interaction as synchronous, creating bottlenecks that block independent agent operation.
Benefits of establishing this best practice:
-
Persistence and retry contain transient failures within the messaging layer instead of exposing them as agent outages.
-
Dead-letter handling helps prevent poison messages from blocking healthy workflow execution.
-
A durable messaging substrate is the foundation for advanced orchestration patterns including saga and arbiter.
Level of risk exposed if this best practice is not established: High
Implementation guidance
Every agent-to-agent call is a coupling decision. Synchronous calls tie the caller's availability to the callee's availability. In a network of agents that multiplies quickly. Five agents with four synchronous dependencies, and the availability product drops below any single agent's SLA. A messaging layer breaks the coupling by buffering the call in durable infrastructure. The caller emits a message and moves on. The receiver processes it on its own schedule, with retries and dead-letter routing handled outside the agent's own code.
Pattern selection follows the interaction shape. Use Amazon EventBridge for content-based routing where a single event fans out to multiple consumers, with EventBridge Schema Registry documenting the contract between agents. Use Amazon SQS for durable point-to-point delivery with configurable visibility timeouts and dead-letter queues. Use Amazon SNS for fan-out to multiple downstream consumers.
Workflow durability ties the messaging layer to business outcomes. A message that reaches its queue still needs orchestration to coordinate multi-step work across agents. AWS Step Functions persists execution state at every step transition, so recovery starts from the last completed step rather than the beginning. Without that persistence, a failure in step five of a seven-step workflow re-executes every prior step, wasting compute and risking duplicate side effects. Dead-letter handling complements durability. Poison messages get isolated for triage rather than blocking healthy traffic behind them.
Implementation steps
-
Map every agent communication path and classify it: Document each interaction as synchronous direct communication (A2A), loosely coupled tool invocation (MCP), or asynchronous event-driven through Amazon EventBridge.
-
Configure EventBridge rules and SQS queues: Set up Amazon EventBridge content-based routing for event-driven paths and Amazon SQS queues for durable point-to-point messaging.
-
Define event schemas in EventBridge Schema Registry: Register a schema for each agent message type so sender and receiver agree on the contract.
-
Configure dead-letter queues with automated triage: Route repeatedly failed messages to DLQs and wire Amazon CloudWatch alarms so operators see poison messages before they block traffic.
-
Instrument the messaging layer with AgentCore Observability: Enable Amazon Bedrock AgentCore Observability for distributed tracing so you can follow a message across EventBridge, SQS, and agent boundaries.
Resources
Related best practices:
Related documents:
Related services:
