---
title: "Conversations"
description: "Review your agent's past conversations and customer feedback to find gaps in its knowledge and improve its answers."
---

Reading real conversations is the fastest way to improve your agent. In your agent, go to **Activity** > **Logs** to see every conversation on every channel. See [Managing Conversations](/docs/human-handoff/managing-conversations) for filters, exports, and deleting.

Each agent reply shows its [action steps](/docs/actions/action-steps) and the widget the visitor saw. Select the summary line to see what ran, what it sent and received, and why other actions did not run. Widgets are read-only and labeled **Shown to the visitor** or **Dismissed by the visitor**.

## A simple review routine

Once a week, or after a busy period:

1. **Start with feedback**

    In **Logs**, filter by **Feedback** and read conversations where visitors marked a reply **Not helpful**.

2. **Find the cause**

    For each poor answer, decide what went wrong: missing information, outdated information, or the agent misunderstood its role. Under each reply, **Read N sources** lists the knowledge it used; expand it to check whether the right pages were found.

3. **Fix it**

    - Missing information: add a [knowledge source](/docs/knowledge) or a [Q&A](/docs/knowledge/questions-and-answers).
    - Outdated information: [retrain](/docs/knowledge/retraining) the source.
    - Wrong behavior or tone: update the [instructions](/docs/agents/instructions).

4. **Check the fix**

    Ask the same question in the [Playground](/docs/agents/playground).

## What else to look for

- **Repeated questions**: if many visitors ask the same thing, add it as a [suggested prompt](/docs/deployments/web-widget/behavior#suggested-prompts) or make it clearer on your website.
- **Missed leads**: visitors who showed interest but weren't asked for their details. Adjust [Collect Leads](/docs/actions/collect-leads).
- **Missed handoffs**: visitors who wanted a person. Adjust [Escalate to Human](/docs/actions/escalate-to-human).
- **Actions that didn't run, or failed**: read the reply's [action steps](/docs/actions/action-steps) in Logs, or try the same message in the [Playground](/docs/agents/playground): what ran, what it sent and received, and why other actions didn't run.
- **Spam or abuse**: tighten the [rate limit](/docs/agents/security#rate-limit) or [allowed domains](/docs/deployments/web-widget/behavior#allowed-embed-domains).