Building ambient agents with Amazon Bedrock AgentCore: From event-driven signals to human-in-the-loop workflows

AWS has published a technical walkthrough for building ambient agents on Amazon Bedrock AgentCore, describing a pattern in which system events rather than user prompts trigger agent workflows.

The post defines ambient agents as responding to event streams, handling potentially many events in parallel, and pausing for human input only when clarification, approval, or review is needed. It cites LangChain among others as having articulated this paradigm, and draws a distinction from fully automated pipelines such as AWS Step Functions, which the post says can orchestrate workflows but cannot reason through ambiguity or ask clarifying questions.

The reference implementation ships Amazon S3 file upload and scheduled event sources as trigger types. API webhooks and database changes are listed as extension points requiring a new handler Lambda function and a corresponding form field on the Signals page. A single setting on a signal governs behavior: with autoExecute false, the default, a job lands in idle status on the Jobs page for human review; with autoExecute true, the job is enqueued to a worker queue and the agent runs immediately, with a human pulled in only if the agent calls ask_human.

The architecture routes an Amazon S3 s3:ObjectCreated notification to a Signal Processor Lambda, which queries a global secondary index on the ambient-signals table to find matching signals and creates a job record per match. Jobs are enqueued onto Amazon SQS, consumed by a Job Execution Lambda that invokes the agent on AgentCore Runtime, and results are written back to Amazon DynamoDB. A React frontend served from Amazon S3 through Amazon CloudFront polls an API Gateway and Lambda tier.

The human-in-the-loop contract uses one tool, ask_human, and a canonical response envelope whose status is completed, interrupted, or error, with matching fields result, question, or error. Interrupted jobs move to interrupted status with requiresAction set to true and appear on an Interrupted tab. The post describes Notify, Question, Review, and Error prompting conventions as ways the agent writes its question, not separate runtime modes.

Prerequisites include Admin access on a sandbox AWS account, AWS CLI defaulted to us-east-1, AWS CDK v2, Docker, Python 3.11 or later, Node.js 18 or later, and access to Anthropic Claude Sonnet 4.5 in Amazon Bedrock. The post states each agent turn is capped at the Lambda 15-minute timeout, and that switching models is a one-line config change.