GENREL03-BP02 Implement timeout mechanisms on agentic workflows
Implement controls to detect and terminate long-running unexpected workflows.
Desired outcome: When implemented, this best practice improves the reliability of your generative AI workload by freeing resources that might have been consumed by unexpected long-running execution loops.
Benefits of establishing this best practice: Automatically recover from failure - Implementing agent timeouts helps to reduce the likelihood of blocking failures on agentic workflows and executions.
Level of risk exposed if this best practice is not established: High
Implementation guidance
Agentic workflows take action on behalf of a user by making calls to external systems. External systems may themselves perform several time-consuming tasks which the agent is not aware of, resulting in idle agents that could run for an extended period of time. In order to maintain a reliable agentic system, customers should implement controls to manage agentic timeout.
One approach to controlling agentic runtime or lifecycle is to implement runtime timeouts on the external infrastructure. For example, if an agent makes a call to an AWS Lambda function through an Action Group, consider applying a timeout to the corresponding Lambda function. The timeout should be set to include the maximum allowable time needed to complete a process, accounting for additional latency for edge cases such as a Lambda cold start. You may consider rounding this value up to avoid unnecessary early terminations.
Alternatively, customers may consider connecting their agentic workflows to an event system, developing an asynchronous process management architecture. Introducing an asynchronous event system gives users the most flexibility and visibility into agent process lifecycle or flow. By requiring the compute underpinning an Amazon Bedrock Action Group to publish events, workload owners maintain insight into where an agent may encounter stalled flow or process. Consider using events to publish agent updates and take action appropriately to prevent long-running invocations.
Error handling at the agent layer should be transparent to users. When errors occur, communicate clear details about the issue while maintaining system security by avoiding exposure of sensitive internal information. The response should outline specific next steps so that users can complete their tasks independently if the agent remains unavailable. This approach promotes operational resilience while maintaining security best practices, as users receive actionable guidance without compromising system integrity. The error messaging framework should focus on user experience by providing alternative paths to task completion while adhering to the principle of least-privilege in information disclosure.
Implementation steps
-
Create an agent within Amazon Bedrock Agents.
-
Define an action group with one or more pieces of supporting compute infrastructure.
-
Implement control logic over the supporting infrastructure.
-
Timeouts are an effective control mechanism. Implement time-outs at the agent layer to terminate a session waiting for a prompt from a user.
-
Alternatively, timeouts can be implemented externally. AWS Lambda action groups can be configured with timeouts and dead letter queues to halt the execution of an external process.
-
Explore asynchronous event publishing to complement timeouts and other control mechanisms.
-
-
Develop recovery logic for the agent to prevent the build-up of half-completed executions.
Resources
Related practices:
Related guides, videos, and documentation:
Related examples:
