Agent Dostoevsky. Homeworkenomics. And a $100 token bill.

After getting our autonomous OpenClaw agent ("Dostoevsky") running on a $125 Raspberry Pi in a previous episode, Jared needed to do market research plan for a 3D-printed hobby rocket business for 5th period Econ, and he thought–let’s use agent Dostoevsky!

Also, I thought, could we do it for free? Why pay LLM token rates for sentiment analysis (a task needed for said business plan) when you can run a free, specialized model?

So we loaded DistilBERT from Hugging Face straight onto the Pi.

We had to load a transformer: $0 (Thank you, open source)

Then the Model: $0 (Thank you, open source)

Compute Environment: $125 (Running locally on the Raspberry Pi, the CAPEX investment funded by me, thank you, Dad)

Trickle-up tokenomics! It felt like a masterclass in parsimony.

Then came the data collection phase.

You need data on which to run the model? Right, of course.

The data our agent found on Hugging Face had nothing about hobby rockets, so we thought about crowd sourcing. Could we use X, Amazon, or Reddit? Trying to avoid any fees we sent Dostoevsky out to scrape comments from YouTube on rocket videos.

I remember a time before YouTube. I remember a time before the internet. I remember that Bill Gates once called linux a “cancer.”

Uh oh. Dostoevsky is not responding.

The local model execution was free—but our agent just spent ~$100 in LLM tokens just planning, retrying HTTP requests, parsing unstructured HTML, and figuring out how to extract the raw text from Youtube.

Go put more tokens in the machine (our agent is from OpenClaw, running on Claude).

We essentially hired a $200/hr executive to do manual data entry!

Dear CloudOps: the hidden cost wasn't the specialized AI work—it's the orchestration loop around it. If you let an LLM handle raw scraping and unstructured data transport, your token budget will bleed out before your model ever runs.

We could have patched the pipeline with a deterministic Python script to pull the comments for free, passing raw JSON directly to DistilBERT on the Pi. We learned this through triage. Total cost moving forward: $0.

How is your team balancing autonomous agent flexibility against token orchestration costs?

PS: If you have happen to need it (and in the spirit of opensource) here is the output:

Hobby-Rocket Market Sentiment (real YouTube data, DistilBERT):

Rank / Segment / Comments / Avg / Sent / % Positive

1 Hobbyist enthusiasts / 92 / +0.245 60.9%

2 Beginners / first-time /87 / +0.174 58.6%

3 Educational / STEM (schools) / 88 / -0.026 46.6%

4 3D-printed rockets / 18 / -0.114 / 44.4%