How it works
How research works
An agent answers one question per run. It plans first, works with on-chain data through tools, and files a report that code checks before anyone sees it.
1. The plan
The first thing an agent does is write a plan: the steps it will take, and a definition for every vague word in the question. "Runner" might become "a token whose peak price reached at least 5x its first trade". The plan is shown on the run and in the report, and the second look checks that the agent stuck to it.
2. Tools
Agents don't browse or guess. They call tools, and every call is recorded with a one-line reason.
| Tool | What it returns |
|---|---|
list_launches | Every token launched on Robinhood Chain in a window, from Pons, pools.trade and new Uniswap v3 pools. |
add_market | Market cap, depth (dollars needed to move the price 2%), price; ETH-quoted markets only. |
add_activity | Trades and volume over the last hours. |
add_price_history | Peak and current price as multiples of the first trade of at least 0.005 ETH, and the peak market cap. Works on thousands of tokens at once. |
add_liquidity_history | How much of a token's launch liquidity was later removed. |
add_names, classify | Names and symbols, how many tokens share a name, and a category per token from a small model. |
read_token | One token in depth: holders, top-10 share, creator share, burned share, market, liquidity. |
creator_tokens | Every token a wallet created, on any of the launchpads above. |
filter_rows, top_rows, group_stats | Work on datasets: filter, rank, and compare groups against the overall rate. |
search_web | Web results, as untrusted excerpts. |
3. Datasets
Big lists (a few thousand launches) stay on the server as datasets. The agent sees a summary and asks for exactly the numbers it needs: a filter, a top 10, a comparison by group. That keeps runs cheap and means every number in the report came out of a tool, not out of the model's head.
4. The report
The report has an answer, findings that name the steps they come from, tables copied straight from tool results (the model can't retype them), the method, and caveats. Before it's accepted, code checks it:
- every number appears in a tool result, in the agent's own tool arguments, or in your question;
- no advice, predictions, "safe" or "scam" labels, or links;
- addresses come from the tools.
A report that fails goes back to the agent with the problems listed, up to two times.
5. The receipt
Every report carries a SHA-256 hash of all the tool results it was written from, the steps, the model, the time and the cost. The public page shows the same receipt.
Limits today
- Robinhood Chain only.
- Market numbers cover ETH-quoted markets; tokens quoted in USDG or tokenized stocks are left out and the report says how many.
- For plain Uniswap v3 tokens, market cap is pool price times total supply, which overstates thin pools.