
Weekend Showcase Challenge: Daily Adventure Agent
š The Season Finale: Turning an Autonomous Agent Into Something You Can Play
Weekend Showcase Challenge: Daily Adventure Agent
š The Season Finale: Turning an Autonomous Agent Into Something You Can Play
For this showcase weekend, I didn't want to start from zero ā I wanted to take my strongest summer build and actually finish the idea I'd left half-done. Two weekends ago I built Once Upon an AI Adventure, an interactive text-adventure game where Amazon Bedrock writes the story live as you make choices. The following weekend, I evolved that concept into Daily Adventure Agent, an always-on agent that wakes up on its own and writes a brand-new mini-story every day, no player required.
Those were two good builds, but they never actually talked to each other. The daily agent produced passive stories you could only read. The interactive game had no memory of anything ā every playthrough started from a blank slate. This weekend's showcase build closes that gap.


šÆ Vision & What It Does
Daily Adventure Agent now does two things instead of one. On its own, unattended, it still generates a complete mini-story every day ā Monday is a fantasy quest, Tuesday is space exploration, and so on through a genre-per-weekday rotation. That part hasn't changed: the agent runs whether or not anyone is watching.
What's new is Ā
play.pyĀ . When you show up, you're no longer stuck just reading what the agent already wrote ā you can pick up exactly where that day's story left off and keep going, making real choices that Bedrock responds to live, for a few turns, until you reach a genuine ending. The auto-generated story becomes the opening act; you become the author of the sequel. Both the daily story and your playthrough get archived and shown together on the live gallery site, so the site tells the whole story: what the agent made on its own, and what you made together with it.

š ļø How I Built It
The core decision was making Ā
play.pyĀ read from the same archive the unattended agent writes to, rather than building a separate game with its own state. Ā play.pyĀ finds the most recently generated Ā story_*.txtĀ , feeds its full text to Bedrock as context, and asks for a new scene that picks up right after the story's existing ending ā then runs the same choice-loop pattern from the original interactive game, three turns to a real conclusion.The trickiest part was making the frontend understand two different content types living in one archive folder. I extended Ā
build_site()Ā to distinguish Ā story_*.txtĀ files (auto-generated, unattended) from Ā playthrough_*.txtĀ files (human-played extensions) and render each with a different label, rather than treating everything as one undifferentiated list. I also had Ā play.pyĀ trigger the same site-rebuild function Ā agent.pyĀ uses, so a playthrough shows up on the live site immediately, not just on the next scheduled agent run.Nothing about the underlying trigger changed from two weeks ago: the daily generation still runs via a background interval loop rather than Amazon EventBridge Scheduler, since the workshop sandbox's IAM role explicitly denies Ā
scheduler:CreateScheduleĀ ā a constraint I diagnosed and documented in the previous build, and simply carried forward rather than re-solving.āļø AWS Services Used / Architecture Overview
- Amazon Bedrock (Nova Lite) ā generates both the unattended daily stories and the live, choice-driven continuations, via direct Ā
invoke_modelĀcalls.
- A scheduled background loop (AWS-hosted compute inside the Builder Center sandbox) ā the unattended trigger for daily generation.
- GitHub Pages ā serves the agent-generated gallery frontend at a permanent public URL.
Background loop --> Bedrock generates daily story --> archive/story_*.txt | v You run play.py -> loads latest story | v Bedrock generates live continuation, turn by turn | v archive/playthrough_*.txt | v build_site() rebuilds index.html from both | v Pushed to GitHub -> served via GitHub Pages
š What I Learned Across the Summer
The biggest lesson of the whole series, reinforced this weekend: diagnose the actual environment before designing the architecture, not after. Every one of my three AWS Builder Center sandboxes had a different IAM scope ā one allowed Nova but not Lambda creation freely, another allowed only Anthropic models via an inference profile, this one allowed direct Nova access again. Reading real error messages and testing real API calls, rather than assuming a "sandbox" means unrestricted access, saved hours across every single build.
The second lesson was about scope discipline. This showcase build isn't bigger in AWS services ā it's still just Bedrock and a background loop. It's bigger in what it lets you do: the combination of "runs without you" and "responds to you when you're there" is more interesting than either half alone, and that came from connecting two existing pieces rather than adding new infrastructure.
Shoutout to Lewis Sawe, @benfowleraws whose Museum That Grows agent from last week's featured spotlight ā an agent that invents one new artifact a day and adds it to a permanent collection ā was the direct inspiration for having Daily Adventure Agent treat its archive as something that accumulates and is worth showing in full, not just overwriting the latest entry.
š Final Thoughts
A creative app, then an agent, then a showcase that puts the player back in the story the agent already started. That's the arc of the summer for this project ā and probably the most honest kind of "finishing touch": not a new feature bolted on, but two things I already built finally working together.
#application #challenge #AWS #GenerativeAI #AmazonBedrock #AWSBuilderCenter
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article