> ## Documentation Index
> Fetch the complete documentation index at: https://docs.poly.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Improve your agent after a real call
> Read the transcript, find the failure, ask Studio Assistant to fix it. The loop that gets the agent from working to actually good.
This page covers the iterate-after-a-call loop with [Studio Assistant](/studio-assistant/introduction): read the transcript, ask Studio Assistant to fix it, share the link. Studio Assistant is available to **project admins on enterprise clusters** as well as free-trial accounts — the workflow is the same in both.
Real callers say things your prompts did not anticipate, mispronounce reference numbers, and go off-script in ways no demo covers. The fastest way to improve an agent is to read the transcripts, identify the failure, ask Studio Assistant to fix it, and share the resulting build for review.
## Quick reference
| Issue | Where to look | What to do |
| --------------------------------------------- | ----------------------------- | --------------------------------------------------------------------------------------- |
| Agent gave a wrong answer | Conversations → transcript | Tighten the topic that matched. Add the right sample questions to the correct topic. |
| Agent missed a topic that should have matched | Conversations → Diagnosis tab | Open the topic, add 3-5 phrasings closer to how the caller actually asked. |
| Agent went off-topic | Conversations → transcript | Add a behavior rule in **Behavior > General > Behavior**, or tighten the topic content. |
| Agent collected info but didn't act on it | Conversations → Diagnosis tab | Wire up the action on the topic, or add a handoff step. |
| Welcome message is wrong | Studio Assistant | *"Rewrite the welcome message to be warmer and mention reservations specifically."* |
| Voice sounds wrong | **Voice > Agent** | Pick another voice. Test on the next call. |
## The loop
### 1. Find the call
Open **Conversations**. Every call lands here, including your test calls. Sort by date, filter by environment, or scan the PolyScore column for the lowest ratings.
Click into one. The right panel opens: transcript on the left side, debugging tabs (Transcription, Scores, Details, Custom metrics) on the right.
### 2. Read what happened
Look for:
* Where the caller pushed back, repeated themselves, or asked the same thing twice. That's usually a wrong-answer or missed-topic moment.
* Where the agent went quiet, said *"I'm sorry, I didn't understand"*, or asked for the same piece of info more than once. Likely a flow validation issue.
* Where the agent answered something it shouldn't have. Guardrail or topic-scoping problem.
* What the caller wanted that you didn't expect. Sometimes the failure is missing intent coverage, not broken logic.
The Diagnosis tab tells you why. It shows which topics matched, which flows ran, which functions were called, and the LLM input on each turn.
### 3. Ask Studio Assistant to fix it
Be specific. The plan you get back is only as good as what you said:
* *"The agent told the caller we offer same-day delivery. We don't. Update the shipping topic to say next-day at best, and add a confirmation step before quoting any delivery time."*
* *"On the booking flow, the agent kept asking for the date even after the caller said 'tomorrow'. Fix the entity extraction for relative dates."*
* *"Add a new topic for 'where is my refund' that says we process refunds in 5-10 business days back to the original payment method."*
* *"The welcome message is too formal. Make it warmer, like a friendly receptionist."*
Studio Assistant proposes a plan. Read it. Approve when it looks right, or send feedback if it doesn't.
### 4. Test the fix
Open the agent in the browser (**Test** in the top-right) and replay the scenario that broke. If it works, promote the new version from Sandbox to Live. If not, send Studio Assistant another revision; the branch is still open.
### 5. Send the link to someone new
You know what you fixed, so you're biased. Send the [share link](/widgets/share) to someone who isn't in the loop and ask them to try the thing you broke. They'll find the next thing.
## One change at a time
Resist the urge to batch fixes. One change per request is easier to plan, easier to review, easier to undo. The fair-use allowance favors small focused requests too.
## Related
Where every call goes after it happens.
Patterns for getting good fixes from short prompts.
The knowledge layer most fixes touch.
Three routes for getting it in front of testers.