AWS Builder Center
Weekend Creative Agent Challenge: Windowseat

Weekend Creative Agent Challenge: Windowseat

Windowseat is an always-on AI creative agent that explores a new real-world location every day, using its actual weather and local time to create an illustrated fictional diary entry. It remembers where it has been and what it has created, ensuring each day's story is different and ready to discover when you return.

AWS Community Builder
What is happening, right now, in a place you have never been?
That question is the seed of Windowseat — an always-on creative agent that wakes up once a day, picks a real coordinate somewhere on Earth, pulls the actual live weather and local time for that spot, and writes a short “diary entry from a stranger's window” along with an illustration of the view. No prompt, no user, no button. The agent simply decides where in the world to look today, and leaves the result waiting in a gallery.

Vision: What It Does

Most daily-art agents invent a universe from nothing — a random prompt spun into a random image. Windowseat takes a different approach: it grounds its imagination in real, verifiable data. Every entry is tied to a genuine place, a genuine moment, and genuine weather conditions, but the “diary” written about that moment is fiction — a small voice imagining what it's like to be looking out a window there right now.
One day the agent might land on a fishing village in northern Japan at 6 a.m. in light snow, and write a two-paragraph diary entry from a lighthouse keeper watching the boats leave. The next day it might land on a rooftop in Lagos at dusk during a thunderstorm, imagined through the eyes of a street vendor packing up early. The illustration always matches the real weather and time of day — so the creative output is anchored to something true, even while the story around it is invented.
The problem this solves: creative tools are reactive by default. Windowseat instead treats “noticing the world” as a background process — something that happens whether or not anyone is watching.

How I Built It

The pipeline runs on a daily schedule, not a user click.
  • Amazon EventBridge Scheduler fires once a day, invoking an AWS Lambda function.
  • Lambda checks Amazon DynamoDB, which stores every coordinate the agent has already visited, the style/tone used, and a rolling diversity score — so the agent avoids revisiting the same regions or repeating the same emotional tone two days in a row.
  • Using that memory, Lambda selects a new coordinate, biased toward landmasses it hasn't explored recently — a simple geographic-spread algorithm rather than pure randomness. This was the part of the project I enjoyed most, since it borrows directly from my day job working on GIS/mapping platforms as a QA engineer.
  • Lambda calls a public weather API for that coordinate to get real conditions and local time, then passes everything — coordinate, weather, time of day, recent themes to avoid — to a text model on Amazon Bedrock, which drafts the diary entry and a visual description.
  • That visual description becomes the prompt for an image-generation model (Amazon Nova Canvas) on Bedrock, producing the illustration.
  • The finished postcard (image + text + metadata) is stored in Amazon S3, and the day's coordinate/style is written back to DynamoDB so tomorrow's entry stays distinct.
  • A lightweight static site reads the latest entry from S3 and displays it — no server needed between generations.
The biggest challenge was making the randomness feel intentional rather than chaotic. A pure random lat/long generator mostly lands in the ocean or produces emotionally flat entries. I had to build simple guardrails — landmass biasing and a memory-based diversity check — so the agent's choices read as curious rather than arbitrary.

Tech Stack

AWS Service
Role in Windowseat
Why It Was Used
Amazon EventBridge Scheduler
Daily autonomous trigger
Wakes the agent on a schedule — no server waiting around, no user click needed
AWS Lambda
Agent orchestrator
Coordinates every step of the pipeline; serverless and idle between runs
Amazon Bedrock (Text Model)
Writes the diary entry
Turns coordinate + weather + memory into a short narrative
Amazon Bedrock (Nova Canvas)
Generates the illustration
Produces the postcard image matching the real weather/time
Amazon DynamoDB
Agent memory
Tracks visited coordinates, tones, and diversity so entries don't repeat
Amazon S3
Postcard & gallery storage
Stores the finished image + metadata; feeds the website
IAM
Access control
Scopes the Lambda role tightly to only the Bedrock, S3, and DynamoDB actions it needs







Architecture Overview

The system follows a fully event-driven, serverless pattern: nothing runs continuously, and no server sits idle waiting for the next day. Amazon EventBridge Scheduler is the only thing that ever “wakes up” on its own, and it does so once a day, invoking the AWS Lambda function that orchestrates the entire run.
Lambda first reads from Amazon DynamoDB to pull the agent's memory — every coordinate it has already visited, the tones and styles it has used, and a diversity score. That memory is what lets the agent behave less like a random generator and more like something with a sense of where it's already been. Using that context, Lambda asks Amazon Bedrock's text model to pick a new coordinate, look up the real weather and local time for it, and draft the diary entry. The resulting visual description is then passed to Bedrock's image model (Nova Canvas) to generate an illustration that matches the real conditions at that place and moment.
The two outputs — story and image — are combined with metadata and written to Amazon S3, which acts as both the permanent archive and the source the website reads from. DynamoDB is updated in the same run so tomorrow's decision has today's choice to react to. IAM keeps the whole pipeline locked down: the Lambda execution role can only call the specific Bedrock models, read/write the specific DynamoDB table, and write to the specific S3 bucket it needs — nothing more.
Because every piece of this only runs for the few seconds it takes to generate a single postcard, the entire architecture comfortably fits inside AWS Free Tier limits, and there is no infrastructure to manage between one day's postcard and the next.
Architecture flow: Schedule → Lambda → Weather lookup → Bedrock (text) → Bedrock (image) → S3 → Gallery, with DynamoDB feeding the agent's memory at every run.
Windowseat architecture — Schedule → Lambda → Bedrock (text + image) → S3 → Web, with DynamoDB memory feeding the agent's decisions.


What I Learned

Coming from a QA background, I'm used to thinking in terms of expected vs. actual output — so the most interesting part of this project was designing an agent whose “expected output” is intentionally never the same twice. I learned that grounding a generative agent in one real, external fact (actual weather, actual time) is often more powerful than giving it total creative freedom — the constraint is what makes the output feel alive rather than random. I also came away with a much better feel for how EventBridge, Lambda, and Bedrock combine into a genuinely serverless agent loop that costs almost nothing to keep running.

Try It​

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