The promises and the honest questions
Why not just ask a chat tool?
Why it works this way
Discover Agents is built around the test you set up first, so what people tell you is filed against the question you are testing.
A test is set up before there is any evidence to bend it. You name your question before the first conversation, and on a test of one idea you write the bar too, the pass or fail line. Once evidence exists, the bar is locked. So a result is harder to bend to flatter your idea: what people tell you is read against the question you set, not against whatever sounds interesting afterwards.
What we gave up: The freedom to ask anything in any order. A test has a shape, and the shape asks more of you at the start.
What it is
It is a fair question, and sometimes the fair answer is: do. If you want a quick read of one transcript, or help wording a message, a chat tool may be all you need.
This guide is about a different job: finding out whether an agent is worth building before you build it, over weeks, with real people. It says what Discover Agents does for that job and why. It does not rank Discover Agents against any other tool.
How to use it
- Decide which job you have. A quick read of one conversation is one job. A test you will run, record and decide on is another.
- For the test, start here. Name what you are comparing, your question, your bar and the people you will talk to. Then add the real conversations as they happen.
- Keep using whatever else you like for drafting and thinking. Keep the conversations, the quotes you rely on and the decision on your board.
What comes back is filed against that test:
- On a test that compares options, each finding lands on one of your cards, or is kept apart as being about none of them.
- A quote saved to your board from one of your conversations is checked against the transcript you added. If the words are not there, it is not saved as a quote, and Iris, your AI guide, tells you it was left out.
- Simulated people are AI, marked "Simulated", and nothing they say counts toward your test.
- A finished test is a record. It opens on what it found, dated and counted.
What it will never do
- Tell you a pass means your idea works. A pass means the next tests are worth running.
- Count a simulated person as evidence.
- Choose your road for you. At the end of a test, nothing is picked for you.
If something looks off
- You only have one transcript and a quick question. Use whatever is fastest. Come back when you are ready to run a test.
- You pasted notes, not a transcript. When what you add reads as your notes rather than the conversation, nobody in it is quoted as having said something.
- It asked for more up front than you expected. That is the trade. The question and the bar come first so the evidence cannot decide them afterwards.