AWS Builder Center

Weekend Creative Agent Challenge: Daily Cloud Haiku

An always-on AWS agent that writes a new cloud/DevOps haiku every day — no manual trigger. Built with EventBridge Scheduler, Lambda, and Amazon Bedrock (DeepSeek V3.2), styled by real-time Lahore weather and saved to S3.

Cloud Architect

Vision & What It Does

As a DevSecOps engineer, I spend most of my week reading logs, alerts, and dashboards full of the same recurring cast of characters — CI/CD pipelines, IAM policies, autoscaling groups, database failovers. Daily Cloud Haiku takes that world and turns it into something quieter and more reflective: a small, always-on agent that writes one original haiku about a cloud or DevOps concept every day, styled subtly by that day's real weather in Lahore.
The idea behind the challenge prompt — "the best tool is the one you never have to open" — fit perfectly with something I already think about a lot in my day job: automation that runs unattended and simply has something ready for you when you check in. Daily Cloud Haiku doesn't need me to open an app, click a button, or feed it a prompt. It wakes up on its own schedule, picks a random cloud topic, checks the weather, writes a haiku, and quietly publishes it to a gallery page. All I do is visit the page later and see what it made.
It's a small piece of software, but it's a genuine always-on creative agent: no manual trigger, no human in the loop, just a scheduled process that produces something new every day and leaves it waiting for me.

How You Built It

I built this over a weekend using services I already had some familiarity with from an unrelated Bedrock-based project, which made the AWS side faster to get right. The core loop is simple: an EventBridge Scheduler cron trigger wakes up an AWS Lambda function once a day. The Lambda function calls a free weather API (wttr.in) to get Lahore's current conditions, picks a random cloud/DevOps topic from a fixed list, and sends both into a prompt for a Bedrock-hosted language model (DeepSeek V3.2) asking for a single 5-7-5 haiku. The result gets written to S3 as a dated JSON file, and a rolling  latest.json  file is updated so a static gallery page can always show the most recent entries.
The biggest early obstacle was simply the model ID. Amazon Bedrock's newer console groups third-party models (DeepSeek, Mistral, Qwen, and others) into one large catalog, and the ID string I initially guessed for the model didn't match what Bedrock actually expected — I got a  ValidationException: model identifier is invalid  error on my first test invoke. The fix was going into the model's own detail page in the Bedrock console and copying the exact  Model ID  field rather than guessing the naming convention.
I originally planned to pair each haiku with an AI-generated image using Nova Canvas, but that model wasn't available in my account without a separate access request, and rather than lose a day of the weekend waiting on approval, I simplified the agent to a text-only creative output — which the challenge prompt explicitly allows ("words, images, sound, or other output"). That turned out to be the right call: it kept the whole build small enough to finish, test, and schedule inside a single weekend.

AWS Services Used / Architecture Overview

  • Amazon EventBridge Scheduler — daily cron trigger (Asia/Karachi timezone) that invokes the Lambda function with no manual step involved
  • AWS Lambda (Python 3.12) — orchestrates the whole flow: weather fetch → prompt construction → Bedrock call → S3 writes
  • Amazon Bedrock (DeepSeek V3.2) — generates the haiku text from a topic + weather prompt
  • Amazon S3 — stores each day's haiku as a dated JSON file, maintains a rolling  latest.json , and hosts the static gallery page via S3 static website hosting
  • AWS IAM — a scoped execution role grants the Lambda function only the permissions it needs (Bedrock invoke, S3 read/write, CloudWatch logging)
  • Amazon CloudWatch Logs — every run's success or failure is logged, which is also how I verified the agent was actually running autonomously rather than just working when I tested it manually
Flow:  EventBridge Scheduler → Lambda → (wttr.in + Bedrock) → S3 → static gallery page 

What You Learned

The most useful lesson was a small one: don't trust a guessed model ID on a fast-moving model catalog — always copy it straight from the console. It cost me one failed test run, but it also taught me to read the Bedrock model detail page more carefully, which will save time on future Bedrock projects.
The second lesson was about scope. I went in assuming an image model was a required part of "creative output," and when Nova Canvas wasn't available in my account, I initially treated that as a blocker. Re-reading the actual challenge prompt reminded me that text is just as valid a creative medium as an image, and cutting the image step turned a stalled weekend into a finished, working agent. Shipping a smaller, fully working thing beat holding out for a larger, half-finished one.
Finally, watching CloudWatch Logs fill up with independent, unattended runs — each with a different topic and a different haiku — was a small but genuine reminder of why "always-on" agents are interesting: the value isn't in any single output, it's in not having to ask for it.

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