AWS Builder Center
Weekend Creative Agent Challenge: Muse Daily

Weekend Creative Agent Challenge: Muse Daily

As a follow-up to last week's creative app challenge. Instead of a tool you open and use, I build an always-on agent: something that generates new creative output on its own schedule and has it ready when you check back, with no manual trigger required. Suggested directions: a daily art/poem generator, output themed to day/weather, remixing community prompts, or a style that evolves over time.

Vision and what it does

Last week I built  Muse , a story/poem generator you had to actively sit down and use, type a theme, press a button, get a piece of writing. This week's challenge asks the obvious next question: what if the app never needed you to open it at all?
Muse Daily is that answer. Every morning at 06:00 UTC, an autonomous agent wakes up, checks the real weather in Helsinki, and writes a brand-new short story or poem themed on that day's conditions , "a mostly clear Friday in Helsinki, 14°C" becomes the seed for something nobody asked for and everybody gets anyway. By the time you open the site, it's already there, dated and waiting, sitting next to a scrolling archive of every previous day's piece. The problem it solves is small but real: the best creative tool is one you don't have to remember to use. The manual "generate one yourself" button from last week is still there for anyone who wants to intervene, but the app's real job now is to keep working while you're not looking.

How I built it

The architecture is a straightforward event-driven pipeline: an EventBridge scheduled rule fires a Lambda function once a day; that Lambda calls the free, key-less Open-Meteo API for current conditions, builds a theme string from the weather code, temperature, and day of week, and calls Amazon Bedrock to generate the piece. The result is written to a single-table DynamoDB design (partition key  "AGENT" , sort key = ISO date) so a  Query  with  ScanIndexForward=False  cheaply returns the most recent entries without a scan. Two more Lambdas expose  GET /today  and  GET /archive  behind API Gateway, and a third keeps last week's manual  POST /generate  alive. The frontend is still a single static HTML page on S3, no framework, no build step, that fetches  /today  and  /archive  on load and renders whatever it finds, including a graceful "the agent hasn't run yet today" state.
The genuinely hard part of this weekend wasn't the agent logic, it was Bedrock itself. Every Converse API call, on both the plain on-demand model ID and a cross-region inference profile, failed instantly with  ThrottlingException: Too many tokens per day , including the very first request ever made with zero prior usage. That ruled out normal exhaustion and pointed at a true account-level entitlement of zero for Nova models' on-demand throughput. I dug through IAM policy history in the account (this AWS account already runs a few other serverless projects, each with its own narrowly-scoped inline policy, a pattern worth copying) and confirmed models that had previously worked, like Nova Micro, were now blocked too, and that the Service Quotas console marked these specific limits as not self-service adjustable. The fix, it turns out, isn't a code problem at all: it requires an AWS Support case to manually raise the account's Bedrock quota. I filed one and kept building everything downstream so the app is fully wired, tested for correct IAM permissions and API shapes end-to-end, and ready to generate real output the moment the quota clears, no redeploy needed.

AWS services used and architecture overview

  • Amazon EventBridge - a scheduled rule ( cron(0 6 * * ? *) ) triggers the agent daily.
  • AWS Lambda - four Python functions: the scheduled agent, manual generate, and two read endpoints.
  • Amazon Bedrock (Nova Lite) - generates the story or poem via the Converse API.
  • Amazon DynamoDB - single-table store for daily entries, on-demand billing.
  • Amazon API Gateway - REST API fronting the three HTTP-exposed Lambdas.
  • Amazon S3 - static website hosting for the frontend.
  • AWS IAM - least-privilege roles per function, scoped to exactly the resources each needs.
EventBridge (daily cron)
→ Agent Lambda → Open-Meteo (weather) → Bedrock Nova (Converse) → DynamoDB
Browser (S3 static site)
→ GET /today, GET /archive → API Gateway → Lambda → DynamoDB
→ POST /generate → API Gateway → Lambda → Bedrock Nova

What I learned

The most useful thing I learned this weekend wasn't a new API, it was how to root-cause an AWS quota problem methodically instead of guessing: reproduce with the smallest possible request, test IAM permission and service quota as two independent variables, try the officially-suggested workaround (cross-region inference profiles), and only escalate to Support once client-side options are genuinely exhausted. I also picked up the single-table DynamoDB pattern for time-ordered data (constant partition key, date as sort key) for the first time in a serverless context, which is a much cheaper way to get "most recent N items" than a scan-and-filter approach. And EventBridge's scheduled Lambda trigger turned out to be almost embarrassingly simple to wire up in SAM, a few lines in the  Events  block of the function, no separate Scheduler role to manage.

Link to app or repo

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