Watch an Agent Think

What you’re looking at

“Agent” is the most overused word in AI right now. Underneath, it’s something simple you can watch happen:

1. The model gets your question, plus a short list of tools it’s allowed to use: here, keyword and meaning-based search over my Japanese deck (733 JLPT words from Kotoba Dash) and my SAT deck (293 words), counting a deck, and a calculator.
2. Instead of answering straight away, it can reply with “call this tool with these arguments.”
3. My server runs the tool, hands back the result, and asks the model again.
4. That loop repeats until the model decides it knows enough to answer.

The page shows every step as it happens. Thinking lines are the model’s own summaries of its reasoning. Tool calls are real searches over real data. For example, ask how long it takes to learn every N5 word at 7 a day, and you’ll see it count the deck (261 N5 words) and calculate ⌈261 ÷ 7⌉ = 38 days, rather than guessing either number.

Why the tools matter

A language model on its own doesn’t know what’s in my decks, and it’s unreliable at arithmetic. With tools it doesn’t have to be: it looks things up and computes them. That’s the difference between an answer that sounds right and one that’s grounded.

Where it went wrong, and the fix

The agent is only as good as its tools. In the first version, asked for “SAT words that describe someone stubborn”, it found obstinate and recalcitrant, whose definitions contain the word “stubborn”. It missed dogmatic, which is in the deck and fits, because its only search tool matched keywords, not meaning. The model even thought about other synonyms, but it could only search for words it guessed might appear in the definitions.

So I gave it a better tool instead of a smarter model. find_words_by_meaning compares meanings rather than letters:

• Every word in both decks was turned into an embedding, a list of numbers that places words with similar meanings close together, by Qwen3-Embedding-0.6B on Cloudflare.
• The embeddings were trimmed to 256 numbers and compressed to one byte each, so both decks fit in about 350 KB that ships with the server code.
• At question time only your phrase is embedded, and the closest words come back with a similarity score. Now “someone stubborn” surfaces dogmatic too.

Meaning search is fuzzy, though: lower-ranked results are sometimes loose matches (daunting shows up for “stubborn”). So the agent is told to check each one before using it, and you can watch it do that in the trace. Most real agent failures look like the original bug: a sensible plan run through a tool that can’t see what it needs to.

Guardrails

Each visitor gets 5 runs a day. The agent is capped at 6 rounds of tool use per question, requests pass a Cloudflare human check, and the lab has a daily spending cap. The agent sticks to studying and the decks; ask it for something else and it will politely say no.