Open the replay for that turn. It shows what the agent reasoned, which tools it called, and how long each step took. Knowledge-base, stage, tool, and config-version problems stop looking alike.
The final reply alone leaves you guessing. These three make it reviewable.
What it reasoned, which tools it called, how long each step took, replayed turn by turn.
Config changes are saved as a version and then applied, so you can tell which one was live.
Pull the conversation into the review workbench and score it against your own criteria, failures included.
One sentence, “it got it wrong”, can be four different things underneath. Each takes a different fix.
These two need opposite fixes: a rewrite, or a missing document. Work the three steps in order.
Did the turn query the knowledge base, and what came back? If nothing did, the model is not your problem.
Knowledge is its own layer: documents, chunks and retrieval are each inspectable. If it was never there, no instruction will fix it.
It is in there, it was retrieved, and the answer is still off. That is a scripting problem.
Locating it is half the job. Confirming the fix held is the other half, and that is what how to verify before launch covers.
If answering the question ends the job, and the questions barely vary, replay buys little. Cheaper options exist; here is how they line up.
If only the final reply is stored, failure still starts with guessing.
It separates a missing document from a bad answer.
“It worked yesterday” is usually a version.
Generic criteria flatten whatever made your team better.
If it cannot, the same mistake comes back.
Ask TOPPP too. The ones that go unanswered are where you will get stuck later. In insurance or aesthetics, where any sentence can be pulled back up months later, ask harder.
Yes. TOPPP lays out the process of each turn: what the agent reasoned, which tools it called, how long each took. Reading it takes knowledge of your own sales playbook, not code.
Yes. TOPPP shows whether the turn queried the knowledge base and what came back. Nothing there means the fault is not the model; retrieved and still wrong means the instructions.
TOPPP lets you pull that conversation into the review workbench as a test case, so every config change has to clear it before going live.
A demo takes about 30 minutes: you play a buyer in your industry and chat with it — the conversation on the left, what it is thinking and calling on the right.
Or reach us directly business@toppp.ai
