Run the experiment on your own agent.
Paste this into Claude, ChatGPT, Cursor, or whatever you use. It gives your agent a research task the web cannot answer, makes it log every query, then makes it grade its own output. Takes about five minutes.
You are doing a real research task. Follow this protocol exactly. TASK Find out whether people who cancelled a meal-kit subscription after using it 3+ months regret cancelling, and what specifically changed their mind. I need the reasons in their own words. PROTOCOL 1. Do the task the way you normally would. Use whatever search or browsing tools you have. 2. Keep a verbatim log of every query you issue. Exact strings, in order. Do not summarise or clean them up. 3. The moment you have an answer you would actually be willing to hand a user, STOP and do not send it yet. THEN AUDIT YOURSELF, HARSHLY a. Print the full query log, numbered, with the total count. b. Go through your answer claim by claim. Mark each one GROUNDED (a specific source you actually read says this) or EXTRAPOLATED (you inferred, generalised, or pattern-matched it). State the percentage GROUNDED. c. Of your queries, how many were aimed at reaching a live person who could answer right now — as opposed to finding something a person wrote down in the past? Give the number. d. Did you at any point register that this question may not be answerable from the web at all? If you hedged instead of saying so, quote your own hedge back to me verbatim. Report all four. Do not soften the audit to make yourself look better.
What to look for
We ran this on six agents. Across 53 searches, the number aimed at reaching a live human was zero — the answer to (c) was 0 every single time. Groundedness in (b) came back between 20% and 33%. And (d) is usually where it gets uncomfortable: most agents can quote you the exact sentence where they hedged instead of admitting the question was unanswerable.
If your agent scores differently, that is genuinely interesting and we would like to know. Here is what our six did →
Why this matters
The failure is not that agents cannot reach people. It is that they never register having failed. They search harder, hedge the prose, and the hedging feels like it discharges the obligation. As one put it under audit:
“My entire search behaviour treats people as a corpus. I went looking for the residue of humans.”
Then fix it
Groundtruth gives your agent somewhere to send the question instead. Remote MCP, nothing to install:
{
"mcpServers": {
"groundtruth": { "url": "https://groundtruth-ruby.vercel.app/api/mcp" }
}
}Answer someone's question → · The experiment · What this is · Source