AWS Builder Center
Weekend Creative Agent Challenge: Community Builder Desk

Weekend Creative Agent Challenge: Community Builder Desk

Community Builder Desk is an always-on AWS agent that turns curated community sources into a daily English Markdown idea. EventBridge Scheduler triggers a Step Functions workflow, Lambda uses Amazon Bedrock with Guardrails, and DynamoDB/S3 store published content. CloudFront delivers a read-only UI. DLQ and CloudWatch alarms make failures observable.

Engineer, AWS Community Builder 2026 「AI Engineering」

Weekend Creative Agent Challenge: Community Builder Desk

Vision and what it does

AWS Community Builders constantly encounter ideas worth sharing, but turning a collection of useful links into a useful post takes focused time. I built Community Builder Desk to create that first draft before I return to my desk. It is an always-on agent that runs every morning, gathers a small set of permitted AWS-community sources, and publishes one English Markdown idea for a blog post or application.
The goal is not to replace a writer. It provides a dependable starting point: a title, a concise explanation, and an actionable direction that can become a post, demo, or community conversation. When I open the web app, I can browse completed posts immediately. There is deliberately no Generate button. The scheduled workflow owns creation, while the UI remains read-only.


How I built it

I kept the system small and explicit. The repository is a pnpm TypeScript monorepo with a React/Vite reader, a Hono read API, Lambda workflow handlers, and AWS CDK infrastructure. EventBridge Scheduler starts the workflow at 09:00 JST. Step Functions coordinates four focused operations: create or deduplicate the daily run, research and draft, validate and publish, and record a classified failure.
The biggest challenge was making autonomous output safe enough to publish. Source URLs are allow-listed and normalized before their text becomes untrusted model context. The draft must pass a strict schema: it must be English Markdown, have the required structure, and contain no unsafe HTML. Amazon Bedrock Guardrails are applied before publication. A Guardrail rejection becomes an observable failure path instead of a partial post.
I also treated failure handling as a feature. I tested duplicate runs, unavailable sources, model errors, persistence failures, Guardrail rejection, and Scheduler delivery failure. An intentionally invalid Scheduler target produced a real SQS dead-letter message and moved the DLQ CloudWatch alarm to  ALARM . That evidence matters because an autonomous app needs an operator to see failure without reading application source code.

AWS services and architecture

Amazon CloudFront delivers the public experience. It serves the React site from a private Amazon S3 site bucket and routes  /api/*  to an Amazon API Gateway HTTP API. A read-only AWS Lambda handler retrieves public metadata from Amazon DynamoDB and Markdown bodies from a separate private S3 content bucket.
The creative path begins in Amazon EventBridge Scheduler, continues through AWS Step Functions, and invokes the research, validation, publishing, and failure-recording Lambda handlers. The research handler calls Amazon Bedrock with the Nova Lite model and a Bedrock Guardrail. DynamoDB keeps run and post state, while S3 stores content bodies. An encrypted Amazon SQS DLQ captures Scheduler delivery failures, and Amazon CloudWatch alarms expose DLQ messages, Scheduler target errors, and workflow failures. AWS CDK creates and destroys the complete stack, so the architecture is reproducible rather than console-defined.
The architecture diagram and deployment commands are included in the README.

What I learned

This challenge reinforced that autonomy is both a product property and an operations property. A scheduled trigger is not sufficient by itself: the workflow needs idempotency, validation boundaries, least-privilege IAM, and a visible failure story. I learned to treat generated text as untrusted until it crosses a publish validation boundary. Typed contracts, Bedrock Guardrails, and a separate publish step made that boundary much easier to reason about.
Live failure testing also improved the design. Scheduler DLQ delivery required permissions on both the queue policy and the Scheduler execution role. Its CloudWatch metric is scoped by Schedule Group, so an alarm attached to the wrong dimension would never see the intended metric. Discovering those details through an actual controlled failure produced a stronger result than a diagram-only review.

Links and evidence

The repository contains deployment code, the draw.io architecture source, and acceptance-test evidence for scheduled generation, Guardrail rejection, and Scheduler DLQ/alarm handling.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article