AWS Builder Center
Weekend Creative Agent Challenge: Postales del Ecuador

Weekend Creative Agent Challenge: Postales del Ecuador

An always-on agent that wakes on a schedule, picks a place in Ecuador it hasn't visited, reads the real weather at those coordinates, and writes a short literary postcard from there. Nobody asks it to — when you come back, it's already written.

Last week's challenge was about building a creative app. This week the prompt changed one word and the whole thing became harder: set it free. Turn the app into an agent that makes something on its own, so that the best tool is the one you never have to open.
I live in Ecuador, so I pointed the agent at the country I know. Postales del Ecuador is an autonomous agent that wakes up on a schedule, picks a place in Ecuador it hasn't visited recently, looks up the real weather at those coordinates at that exact moment, and writes a short literary postcard from there. Nobody asks it to. When you come back, it's already written.

Vision & What It Does

Most generative tools have the same friction: you open them, you stare at an empty prompt box, you think of something to ask, you wait. The creative burden stays with you. The tool is a tool.
I wanted the opposite shape. An agent with its own reason to write, that goes looking for the world instead of waiting to be asked.
Every run, without any human in the loop, the agent:
  1. Reads its own memory in DynamoDB — where it has already been, and in which voice it already wrote.
  1. Picks a place from a catalogue of 50 real Ecuadorian locations, discarding the 18 most recently visited. Quito, the Cotopaxi páramo, Mompiche's black sand, the flooded lagoons of Cuyabeno, Puerto Ayora in the Galápagos.
  1. Picks a narrative register from 12, discarding the 8 most recently used.
  1. Fetches the real current weather at those coordinates from Open-Meteo.
  1. Translates that weather into physical sensation — this is the important part, and I'll come back to it.
  1. Writes the postcard with Amazon Bedrock.
  1. Validates its own output against six rules and retries if it broke any of them.
  1. Saves to memory and republishes the public archive to S3.
The output is genuinely conditioned on reality. A postcard from the Cotopaxi páramo written at −6.5 °C with 86% humidity is not the same postcard the same place produces on a clear afternoon. Here's one the agent wrote this morning, unedited:
Caballos en el silencio blancoEl frío entumece los dedos y la piel se eriza. Todo está un poco húmedo, como si hubiera salido de una lavadora olvidada. La brisa suave lleva el olor a tierra mojada y pasto fresco. Los caballos salvajes se mueven lentos, casi fantasmales, entre la vegetación raquítica del páramo. Sus pasos son silenciosos, apagados por la espesa capa de hierba. El cono nevado del volcán se alza imponente, cubierto por una manta de nubes grises. El cielo parece un lienzo sin terminar, una pintura a medio hacer. Las nubes bajas se adhieren a la piel, como si quisieran envolverte en su abrazo frío y húmedo. El aire está saturado, pesa sobre los hombros. Los sonidos son apagados, lejanos. El lugar respira una quietud profunda, como si esperara algo que no llega.
The public page is deliberately read-only. There is no button, no prompt box, nothing to click. It only displays what the agent already wrote by itself.

How You Built It

The first plan died on day one. The original design had an illustration per postcard from Amazon Nova Canvas. When I actually tried to call it, Bedrock answered that the model is marked Legacy and blocked. I checked  us-east-1 ,  us-west-2  and  us-east-2 : there was no text-to-image model available to me at all — the Stability models present are editing operations that require an input image. I tried having Nova write the illustration directly as SVG code. The result was a giant red triangle over a green rectangle.
So I stopped trying to force the image and rebuilt the project around what the account actually does well: text. That turned out to be the right trade. The project got smaller and much more coherent.
The second problem was that the writing was terrible. My first version passed the raw weather data into the prompt, and the model dutifully recited it back:
"la temperatura es de 25.4 grados que se sienten más frescos por la humedad, un 66%..."
That's a weather report wearing a literary hat. I tried instructing the model not to mention numbers. It kept mentioning numbers.
The fix was to stop asking and start constraining. Now the Lambda translates the measurements into physical sensation in code, before the prompt is ever built:
if t < 8: cuerpo = "frío de páramo, de meter las manos en los bolsillos"
elif t < 14: cuerpo = "fresco, se agradece una chompa"
...
if t >= 24 and h >= 70: extra = " Se suda sin moverse."
The model never sees a number, so it cannot recite one. All it can do is write about how it feels to be there. That single change did more for the output quality than every prompt instruction I wrote combined.
The third problem was repetition. An agent without memory writes the same postcard forever. Before generating, it now queries DynamoDB for its recent history and deliberately excludes what it just used — both places and narrative registers. That memory layer is what separates an agent from a cron job with an API key.
The fourth problem was the model quietly breaking rules. It would sneak the location name into the title, or slip in tourist-brochure words, or run long. So the agent grades itself: a validator checks six rules — no digits, no banned words, a sane length, title under seven words, title must not name the place, title must not leak the tone label — and retries up to three times. The two length numbers are deliberately different: the prompt asks for 95 to 125 words, the validator rejects anything outside 85 to 145. That tolerance band exists so a good postcard that ran thirty words long does not burn a retry. Failed attempts are logged, so the reasons are visible in CloudWatch.
The fifth problem was my own infrastructure, and it was self-inflicted. I built it four ways over the weekend. It started as a 171-line shell script, which worked but wrote the answer to "does this already exist?" five times over. Then CloudFormation with three nested stacks organized by lifecycle, which is exactly what the best practices guide recommends and exactly the wrong call for seven resources. Then a single SAM template. Then modular Bash. Then Terraform, which is where it stayed.
Two moments earned Terraform its place, and neither was about elegance.
The first is that it caught a bug the shell script had been hiding. My Bash version attached the X-Ray managed policy with  || true  on the end of the line, so when the ARN turned out to be wrong the failure was silent —  AWSLambdaBasicExecutionRole  lives under  service-role/  and  AWSXRayDaemonWriteAccess  does not. Tracing still reported  Active , because configuring tracing and having permission to write traces are different things. The function had been running for hours without that permission and nothing said so. Terraform refused to apply and named the ARN that did not exist.
The second is that refactoring stopped meaning rebuilding. Moving seventeen resources into modules changes every resource address, and Terraform would ordinarily read that as seventeen new resources — destroying the old ones, including the table holding every postcard the agent had written.  moved  blocks say "same resource, new address":
moved {
from = aws_dynamodb_table.historial
to = module.almacenamiento.aws_dynamodb_table.historial
}
The plan came back  0 to add, 1 to change, 0 to destroy , and the one change was an HTML checksum. Every earlier migration that weekend had cost me a backup-and-restore cycle.

AWS Services Used / Architecture Overview

Amazon EventBridge Scheduler (cron · America/Guayaquil)
│ wakes the agent — nobody clicks anything
▼
AWS Lambda "postales-del-ecuador" (Python 3.12 · arm64 · no dependencies)
│
├─1─► Amazon DynamoDB read memory: places and tones already used
│ └─ picks a place it hasn't visited, in a voice it hasn't used
│
├─2─► Open-Meteo API real current weather at those coordinates
│ └─ translated to physical sensation IN CODE — the model never sees a number
│
├─3─► Amazon Bedrock amazon.nova-pro-v1:0 · Converse API
│ ↺ validates 6 rules, retries up to 3 times
│
├─4─► Amazon DynamoDB persist today's postcard
└─5─► Amazon S3 republish data/archive.json
│
▼
S3 Static Website Hosting → the public page (read-only)

If it breaks: CloudWatch alarm → SNS. Traces in X-Ray.
Service
Role
Amazon EventBridge Scheduler
The agent's clock. It is what wakes the agent up — no human in the loop.
AWS Lambda
The brain. Orchestrates memory, weather, generation, validation and publishing. Runs in under three seconds on arm64.
Amazon Bedrock (Nova Pro)
The voice. Writes the postcard through the Converse API, conditioned on real context.
Amazon DynamoDB
The memory. Prevents repeating places and registers, with point-in-time recovery.
Amazon S3
The archive and the gallery. Stores the postcards and serves the static site.
Amazon CloudWatch
Logs with 30-day retention, and an alarm on function errors.
Amazon SNS
Where that alarm goes, because nobody reads logs at 6 a.m.
AWS X-Ray
Traces of each run.
AWS IAM
Least-privilege roles. Bedrock access is scoped to one model ARN, not  * .
The infrastructure is Terraform — reusable modules in  infra/modules/ , composed by a single environment in  infra/environments/prod/ , with remote state in S3 using the backend's native  use_lockfile , which as of Terraform 1.10 replaces the DynamoDB lock table this used to need.  default_tags  on the provider tags every resource without a repeated  tags  block. An  archive_file  data source builds the deployment package and feeds its hash to  source_code_hash , so changing the Python is enough to ship it. Even  index.html  is an  aws_s3_object  with  etag = filemd5(...) , which makes the front end one more thing a plan can tell you about.
The code is layered so the parts that decide quality can be tested without AWS. The domain holds the rules — the weather-to-sensation translation, the six validation rules, the memory-based selection — and imports nothing but the standard library. Adapters translate DynamoDB items, Open-Meteo responses and Bedrock output into domain models at the boundary, so no storage detail reaches the rules. The 50 places live in a JSON file and are injected, because they are data, not policy.
The retry loop demonstrates the split better than any diagram: the application insists, the infrastructure calls the model, and the domain decides whether what came back is publishable. Seventeen tests cover the rules and run in ten milliseconds with no network and no credentials.

What You Learned

Check what your account can actually run before you design around it. I lost the first hour of the weekend to a model that was listed but unusable. Listing a model and being able to invoke it are different things.
Constraining beats instructing. "Don't mention numbers" failed every time. Removing the numbers from the model's input worked on the first try. When output quality depends on the model not doing something, the most reliable fix is usually to make it impossible rather than forbidden.
Memory is what makes it an agent. Scheduling alone gives you a cron job. What made this feel autonomous was that it knows what it did yesterday and chooses differently today.
An autonomous agent should grade its own work. Nothing else will. Nobody is watching at 6 a.m., so the validate-and-retry loop is not a nicety — it's the only quality control that exists.
Best practices are conditional, and the condition is usually scale. I read the guidance on organizing stacks by lifecycle, applied it faithfully, and produced three nested stacks for seven resources. The guidance was right; my project was too small for it. Reading a rule is easy — knowing when it applies to you is the actual skill, and I got it wrong before I got it right.
Building the same infrastructure four ways taught me more than any of the four. Writing a teardown script by hand is what finally made  terraform destroy  legible to me — not as a command, but as a thing somebody had to build because the alternative is maintaining deletion order yourself. You can read that in a comparison table and nod. Writing the table by hand is different.
A tool that refuses to run is doing you a favour. My shell script ended a line with  || true  and hid a broken IAM policy ARN for hours. Terraform stopped and named the ARN. The value was not the syntax or the state file — it was that the failure was impossible to ignore.
Autonomy has to be provable, not claimed. Every postcard records which trigger produced it, and the public page shows that badge. Saying an agent is autonomous is easy; showing the timestamps is what makes it credible.

Link to App or Repo

Evidence of autonomous generation

Each entry on the public page carries the trigger that produced it. Entries marked "despertado por el scheduler" were generated by EventBridge Scheduler with no human involved. The CloudWatch logs show the same thing from the other side — and this run happens to catch the validator rejecting a draft and asking for another one:
[agent] waking up. trigger=eventbridge-scheduler local_time=2026-08-22T15:29:33.052552-05:00
[memory] 20 previous postcards in the history
[decision] place=Mitad del Mundo tone=melancolico y contenido
[weather] Mitad del Mundo: 16.4C, llovizna ligera, humidity 85%, wind 10.1km/h
[bedrock] attempt 1 model=amazon.nova-pro-v1:0 tokens_in=643 tokens_out=254
[validation] attempt 1 rejected: length 157
[bedrock] attempt 2 model=amazon.nova-pro-v1:0 tokens_in=643 tokens_out=168
[validation] attempt 2: postcard accepted
[memory] postcard 2026-08-22T15-29-33 saved
[s3] archive republished with 23 postcards
[agent] done. 'Entre dos mundos, la quietud' from Mitad del Mundo. archive: 23 postcards
During the challenge weekend I ran the scheduler on a short interval to build up a visible archive; it now runs once a day at 06:00 Ecuador time.
The agent is still out there, writing postcards from places I haven't been in years.
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