
Weekend Creative Agent Challenge: Driftwatch
An always-on agent that writes dispatches from an imaginary coastline every 15 minutes, and rewrites its own style when it catches itself repeating. Bedrock was blocked account-wide, which turned out to be the best thing that happened to the design.
Weekend Creative Agent Challenge: Driftwatch
Tag:
agents Repo: https://github.com/mohitk-patwari/Driftwatch
Live: http://driftwatch-web-232351199908.s3-website-us-east-1.amazonaws.com
Vision & What It Does
Driftwatch is an unattended observation post. Every fifteen minutes, with nobody involved, it reads the real weather over Bengaluru and files a short dispatch from a coastline that does not exist.
The challenge asked for a tool you never have to open, and I took that literally. There is no input field. There is no button. You cannot ask Driftwatch for anything. It has been running since before you arrived and it will keep running after you close the tab — the site is a window onto work already done, not a request you make.
But an agent that produces the same thing forever is a cron job with better prose. So the second mechanism is the one I actually care about: the agent's writing style drifts based on its own accumulated output.
Four parameters define its voice at any moment — mean sentence length, a measure of ornamentation I call austerity, which of five imagery banks dominates, and a measured repetition score. Before writing anything, each run loads its last twenty dispatches and computes what fraction of its content words it has been overusing. Below a threshold it nudges itself gently. Above it, the agent has decided it is stuck, and it pushes hard: switching to its least-used vocabulary and moving its numeric parameters away from their recent mean.
Those parameters persist between runs. The voice at dispatch 400 is not the voice at dispatch 1, and the site plots the distance travelled.
On the first day of operation this happened on its own:
seq 1 drift 0.00 tidal first run, nothing to differ fromseq 2 drift 0.03 tidal gentle nudgeseq 4 drift 0.16 tidal larger moveseq 5 drift 0.03 tidal gentle againseq 6 drift 0.52 mineral threshold crossed, vocabulary switched
Nothing scheduled that jump at seq 6. The agent noticed five consecutive runs of tidal imagery, judged itself repetitive, and changed its own vocabulary bank.
How You Built It
The first command I ran was not code. It was a check on whether Amazon Bedrock would answer.
It did not.
ValidationException: Operation not allowed , across two model families, with AdministratorAccess attached to my user. This is the third time this account has refused Bedrock, and I have learned to treat the preflight check as non-negotiable — the previous time I discovered it ninety minutes into a build instead of twelve seconds in.So Driftwatch generates its prose in deterministic Python, and I want to argue this made the project better rather than worse.
"The agent's style evolves over time" is an unfalsifiable claim when a language model is doing the writing. You cannot inspect the mechanism, you cannot prove the change, and you cannot reproduce it. With explicit numeric parameters the drift is measurable, chartable and deterministic: the same seq, style and weather always produce the same dispatch. The interesting part of the project stopped being an API call and became something I actually had to design.
The second decision that saved time was removing the API layer entirely. My first sketch had API Gateway in front of a read endpoint. Instead, the Lambda overwrites a static
feed.json in the same S3 bucket that serves the site. The browser makes one same-origin request for a flat file. No CORS configuration, no credentials in the client, no request path that can fail — and on this account, Lambda Function URLs return 403 regardless of resource policy, so an entire category of debugging disappeared.The most valuable thing I did was ordering the build wrong on purpose. Evidence of autonomous operation is wall-clock evidence: you cannot compress it at the end. So before writing a single line of the composer, I deployed a stub that wrote a placeholder row on a fifteen-minute schedule. The loop was live and accumulating history while the interesting code was still being written. By the time the real composer shipped, the schedule had already proven itself.
I ran three Claude Code terminals in parallel against a frozen
CONTRACT.md — one on infrastructure, one on the generation engine, one on the frontend — with strict, non-overlapping directory ownership and the data shapes specified in advance. Writing the contract before launching anything is what stopped three agents inventing three incompatible interfaces.AWS Services Used / Architecture Overview
EventBridge Scheduler --(assumes role)--> Lambda (Python 3.12) | Open-Meteo (keyless) ---(weather)------->| | DynamoDB <--(history + style state)--> | v S3 (feed.json) | v S3 static site (browser)
Service | Role |
|---|---|
Amazon EventBridge Scheduler | fires the agent every 15 minutes, unattended |
AWS Lambda | Python 3.12, standard library plus boto3 only |
Amazon DynamoDB | dispatch history and the persistent style vector |
Amazon S3 | static site hosting and the JSON feed |
Amazon CloudWatch Logs | the autonomy audit trail |
AWS IAM | two roles — one the scheduler assumes, one the function assumes |
Everything sits inside the Free Tier. A single table holds both the dispatches and the agent's evolving state, distinguished by partition key.
The weather call degrades rather than fails. If Open-Meteo times out, the run continues with the label
unobserved and still files a dispatch. An outage at someone else's API must never look like a missed dispatch, because autonomy is the entire premise.Autonomy is verifiable rather than asserted:
$ aws logs tail /aws/lambda/driftwatch-agent --since 1h | grep wrote1:05:24 {"seq": 2, "lexicon": "tidal", "drift": 0.034}1:20:25 {"seq": 3, "lexicon": "tidal", "drift": 0.033}1:35:25 {"seq": 4, "lexicon": "tidal", "drift": 0.162}
Exactly fifteen minutes apart. I triggered none of them.
What You Learned
Test the dependency before you design around it. One CLI call at minute zero redirected the entire architecture before a line of the pipeline existed. The same check the previous time cost me ninety minutes because I ran it late.
Read AWS errors as "something refused this," not as literal descriptions. A
ValidationException for an account-level block. A 403 with no hint that a restriction exists. These describe the wrong problem, consistently.Evidence that depends on wall-clock time must be scheduled first. This was the scheduling insight of the whole build. Deploy the ugly loop on day one and improve the generator while history accumulates underneath it.
Constraints are better design partners than capabilities. I would not have built an inspectable drift mechanism if a hosted model had been available. I would have written a prompt and claimed the style evolved. The blocked service forced a design where the interesting behaviour is something I can actually show you a graph of.
There is a version of this project where the agent calls a model and produces prettier sentences. There is not a version where I could prove it was changing.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article