Built-in patterns
Three patterns are available out of the box and run automatically on conversational runs:- Frustration — the user expresses annoyance, dissatisfaction, or negative sentiment toward the agent.
- Confusion — the user indicates they don’t understand or are lost.
- Task failure reported — the user reports that the agent didn’t complete their request correctly.
Custom patterns
Create your own patterns to monitor behaviors specific to your product. Custom patterns use AI-powered classification against a description and, optionally, labeled examples. Ideas teams typically add:- Competitor mentions
- Requests for missing functionality
- Escalation requests
- Compliance-relevant phrases
- Specific workflow behaviors, for example “user asked to cancel”
- Any domain-specific signal that matters to your team
Creating a pattern
Open Patterns and click Create pattern to open the wizard.1
Define the pattern
- Pattern name — human-readable, up to 50 characters. Example:
User asked for refund. - Pattern ID — kebab-case identifier auto-generated from the name, up to 50 characters. Example:
user-asked-for-refund. - Description — describe what the pattern detects, up to 350 characters. The more specific the description, the better the accuracy.
- When to check — all runs (default), runs that use any of a set of tools, or runs that use all of a set of tools.
- Metadata filters — up to 10 key/value filters restricting scans to matching runs.
2
Label examples (optional)
Improve accuracy by labeling recent runs as positive or negative examples.
- Browse runs from the past 7 days.
- For each run, mark it as matches (positive) or does not match (negative).
- Up to 100 examples per pattern.
- Filter runs by text search, run type, attributes, or metadata.
3
Configure alerts
- Enable Slack alerts — toggle notifications for this pattern.
- Select channels — pick channels from your connected Slack workspace.
- Alert frequency — every detection, threshold-based (for example 5 detections in 60 minutes), or a daily digest.
When you create a pattern, it automatically scans recent historical runs for matches so you see results immediately.
Patterns list
The Patterns page shows every pattern (built-in and custom) with its current detection count. Use the search bar to filter by name.Pattern detail
Open any pattern to see the runs where it was detected. Each row includes:- The detection confidence score
- Evidence explaining why the pattern matched
- A link into the full trace
How patterns fit with the rest of the product
- Failures (Failures) surface recurring runtime errors from the execution itself. Patterns surface behaviors the classifier recognizes from the transcript and metadata.
- Insight Search (Insight Search) can be scoped to a pattern’s runs to investigate root cause.
- Recaps (Recaps) reference detected patterns in the trend and issue sections.
- Slack (Slack) delivers pattern alerts to subscribed channels, and the Slack bot can suggest new patterns based on what it sees.
Limitations
- Patterns run on conversational runs only. Runs without
tcc.conversational: trueare excluded. - Custom patterns are classifier-based. Accuracy depends on the description and labeled examples; use the labeling step for behaviors that are ambiguous from words alone.
- Trigger conditions are
ANDbetween tool filters and metadata filters. There is no OR-of-groups composition today.
Related
Failures
Recurring errors from execution, grouped automatically.
Insight Search
Investigate the runs a pattern is matching.
Slack
Deliver pattern alerts to channels.
Concepts
Runs, sessions, and conversational runs.
