Overview
Launching an agent is only the beginning. Without systematic observation, you cannot tell whether users are getting answers they trust, whether integrations are working correctly, or whether the agent's tone matches what you intended.
Silent failures are the most dangerous kind: a user asks a question, receives a confident but incorrect answer, and leaves without providing any feedback. Unless you review conversations and track the feedback users do leave, those failures accumulate invisibly. Over time they erode trust in the agent and in the team behind it.
qlar gives you three dedicated observation surfaces, each designed for a different type of question:
| Surface | Where | Best for |
|---|---|---|
| Dashboard | Home | Volume, trend, function-tool adoption |
| Feedback Insight | Insight → Feedback | User satisfaction, failure pattern detection |
| Monitoring Logs | Monitoring | Technical errors, integration debugging, audit trail |
Use all three together to build a complete picture of how your agent behaves in production.
The Three Observation Surfaces
Dashboard
The Dashboard answers the question: Is my agent being used, and is usage growing?
Open it every morning to catch overnight anomalies. Five metric cards give you an immediate health snapshot, and the conversation chart lets you drill into any time bucket to read the individual threads. The Function Tools pie chart reveals which tools the AI is actually invoking — and which it is ignoring.
Go here when: you want to check overall activity, spot traffic anomalies, or verify that function tools are being used.
Feedback Insight
The Feedback page answers the question: Are users satisfied with the answers they are receiving?
It surfaces the explicit ratings users leave (👍 / 👎) and lets you drill into every individual negative rating to read the user's reason and the full conversation. A negative rate that climbs week over week is the single most reliable early warning that something in the agent has changed or drifted.
Go here when: you want to measure user satisfaction, identify recurring complaint patterns, or track whether a configuration change improved quality.
Monitoring Logs
The Monitoring page answers the question: What exactly happened, and why?
It exposes the raw log stream from the agent's backend services, filtered by source (Published or Draft), severity, and time window. Every log row links to the conversation it belongs to and opens a structured context panel with the full request payload and error detail from any external API call.
Go here when: you need to diagnose a specific error, validate that a configuration change is working in Draft before publishing, or build an audit trail after an incident.
→ Investigate with Monitoring Logs
Build a Monitoring Cadence
Individual observations are useful; a consistent rhythm makes them powerful. The table below is a starting point — adjust frequency and focus to match your deployment scale and user base.
| Frequency | What to check | Goal |
|---|---|---|
| Daily | Dashboard 24h view · Warning and Error logs | Catch silent failures and traffic anomalies before they accumulate |
| Weekly | Dashboard 30d view · Feedback trend · Function-tool pie chart · Export PDF | Measure week-over-week quality trend and prepare stakeholder update |
| After a config change | Dashboard 72h view · Draft logs before publishing · Feedback score post-publish | Confirm the change had the intended effect and did not introduce regressions |
| Monthly | Dashboard 12m view · Avg Messages per Conversation trend · Negative feedback reasons | Identify long-term drift and plan the next iteration of the agent |
| After an incident | Monitoring logs filtered to incident window · View Context on each Error | Root-cause analysis and documentation of corrective action |
Tip: Export the Dashboard as PDF before and after each significant configuration change. Keeping these snapshots lets you demonstrate improvement to stakeholders with concrete data rather than anecdote.
Close the Improvement Loop
Observation has no value unless it drives action. Use the framework below to move from what you see in qlar to what you change in your agent.
Observe → Diagnose → Act
| What you observe | Likely diagnosis | Where to act in qlar |
|---|---|---|
| Negative feedback rate above 20 % | Agent tone, scope, or accuracy is not meeting user expectations | Revise Behavior (Mission, Personality, Interaction) |
| Avg Messages / Conversation rising over several weeks | Agent is not resolving requests in a single reply; users are rephrasing or following up repeatedly | Improve Knowledge resources and tighten Mission scope |
| A specific function tool rarely appears in the pie chart | AI model is not recognising when the tool is relevant | Rewrite the tool's Name and Description in Custom API settings |
| Error logs spiking after a knowledge upload | Document parsing or chunking issue with the new resource | Check the resource status in Knowledge → Resources |
| Traffic drops sharply on a specific channel | Messenger channel disconnected or token expired | Verify connection in Channels → Messenger settings |
| Users citing "wrong information" in negative feedback reasons | Knowledge base is outdated or the source document has changed | Update or re-upload the relevant resource |
| Critical log from a specific service | Infrastructure-level failure outside your configuration | Export log details and escalate to support |
The Iteration Cycle
Every agent improves through iteration. A practical cadence looks like this:
- Observe — Review Dashboard, Feedback, and Logs on the schedule above.
- Diagnose — Identify the one or two issues with the highest impact on user satisfaction.
- Hypothesise — Form a specific change: "If I add a clearer instruction about response length to Interaction, the 'too long' feedback reason should decrease."
- Act — Make the change in Draft mode.
- Validate — Use Monitoring logs filtered to Draft Agent to confirm no new errors. Test manually in the preview panel.
- Publish — Deploy the Draft to Published.
- Measure — Check Feedback trend and Dashboard metrics over the next 72 hours to 30 days depending on traffic volume.
- Repeat — Return to step 1.
Agents that are monitored and iterated regularly consistently outperform those that are deployed once and left unchanged.
Related Pages
- Read the Dashboard — metric cards, conversation chart, function-tool adoption, and PDF export
- Analyse User Feedback — satisfaction trends, negative feedback patterns, and the improvement workflow
- Investigate with Monitoring Logs — log levels, filters, action buttons, and debug workflow
- Automate with Custom API — configure the function tools whose usage you track in the Dashboard
- Configure Behavior — act on insights from Feedback by revising Mission, Personality, and Interaction
- Improve with Simulation — use simulated conversations to test hypotheses before deploying changes to production