
Weekend Creative Agent Challenge: The Museum That Grows
The Museum That Grows is an always-on agent that curates a museum for a civilization that never existed. Once a day it wakes up, reads everything already in the collection, invents one new artifact that fits, and adds it to the gallery
Vision and what it does
Last week's challenge was a creative app. You gave it input, it gave you output, done. The moment I read this week's brief I noticed the real shift hiding in it. An app forgets you the second you close the tab. An agent remembers, works while you are gone, and has something ready when you return.
So I asked what a creative thing can do that only works if it has memory and time. The answer was accumulation. A poem generator produces the same quality whether you run it today or next month, because each poem stands alone. But a museum is different. A museum is only interesting because the objects share a world. Piece 40 means more because you already saw pieces 1 through 39.
The agent invents a people on its first run: the Tidesworn, a coastal culture that worships a Tide-Mother and works in salted bronze, seaglass, and tidewood. Every day after that it adds one artifact and stays inside that world. A coin becomes an amulet becomes a mariner's shard becomes a harp, and the later plaques reference the earlier catalog numbers. The collection writes its own history.

How I built it
I started from the memory model, not the UI, because memory is the whole idea. The archive is a single JSON file in S3. It holds the civilization and an ordered list of artifacts. That file is the agent's brain.
The prompt has two modes. On an empty archive it casts Nova Pro as a curator and archaeologist and asks it to found the civilization and its first object together. On every later run it feeds the civilization summary plus every existing artifact back into the prompt and asks for exactly one new piece that stays consistent, reuses established materials and motifs, and references an earlier object when it makes sense. The model returns strict JSON: name, invented dating, material, a shape from a fixed set, an emoji, a three-color palette, the plaque text, and a cross-reference line.
There is no image generation. The Canvas model on my account is locked to legacy, same as in my earlier entries, so the artifacts are drawn as procedural SVG in the browser. Each shape has its own vector drawing, colored by the palette the model chose and decorated with small "runes" seeded from a hash of the artifact's name. Deterministic, so an object looks the same every visit, but distinct from every other object. CSS turns the rest into a wall of lit vitrines with brass plaques.
The one real snag was permissions. On the first invocation S3 returned AccessDenied instead of NoSuchKey when the archive did not exist yet, so my "is this the first run" check never fired. The cause was a missing s3:ListBucket permission. Without it, S3 hides the difference between "not there" and "not allowed." Adding ListBucket made the missing object report as a clean 404, and the founding path worked.

AWS services and architecture
- Amazon EventBridge Scheduler fires the agent once a day on a cron. This is what makes it always-on.
- AWS Lambda runs the agent in Python with boto3.
- Amazon Bedrock (Nova Pro) invents each artifact as structured JSON through the Converse API.
- Amazon S3 stores the archive and hosts the static gallery.
- IAM scopes the Lambda role to one model and one object, and gives the scheduler a role that can only invoke this one function.
EventBridge Scheduler (daily cron) | invoke v Lambda (Python) | read museum.json (memory) v Amazon S3 <----------------+ | | | civilization + past | v artifacts | Amazon Bedrock: Nova Pro | | one new artifact | +---> append, write back-+ |Browser <--- S3 static site <---+The archive doubles as memory and as the site's data source. No database, no API layer, no server. The gallery is one HTML file that fetches the JSON and renders it.

What I learned
The line between an app and an agent is memory plus a clock. Adding EventBridge Scheduler and a JSON file in S3 was almost no code, but it changed the whole character of the thing. The output is now a growing body of work instead of a single reply.
Feeding the model its own past output is a simple and powerful pattern. Consistency that would be hard to hand-code just emerges, because the model reads the history and matches it. The prompt is the contract, and the archive is the context.
I also learned an S3 permission detail the hard way. If you rely on catching NoSuchKey to detect a missing object, you also need s3:ListBucket, or you will get AccessDenied instead and misread it.
Link
Come back tomorrow. There will be one more artifact than there is today.

Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article