Behaviors

Behaviors is the semantic view of what your AI product is doing in production. Instead of reading thousands of traces one by one, use Behaviors to see repeated user intents, assistant responses, tool activity, issue clusters, volume, and change over time. Behaviors connects observability to action. Use it to decide what deserves an Improve cycle, which traces to inspect, and which datasets or evaluators should cover the issue.

​

What a Behavior represents

A Behavior is a durable pattern detected from traces. It can describe:

• What users ask for.
• How the assistant responds.
• Which tools or retrieval steps appear.
• Where failures, errors, or policy issues cluster.
• How a pattern changes over time.

Each Behavior has a role, such as User, Assistant, or Tool, so you can distinguish what came from the user, the model, or the system. Behaviors are also scoped to the subject being analyzed, such as an agent, prompt, or tool, so the same label in two parts of a project does not necessarily mean the same underlying production pattern.

​

Snapshots and identity

Adaline analyzes Behaviors in snapshots. A snapshot freezes the state of the latest analysis: which Behaviors were presented, how many spans were processed, which roles were observed, and which alerts were emitted. The Behavior identity persists across snapshots. That lets Adaline show whether a pattern is new, active, dormant, changed, vanished, reactivated, or split instead of treating every analysis run as a fresh list of clusters.

​

Issue signals

A Behavior can carry issue signals from production evidence: Use these signals as triage input. They are evidence, not a final root-cause analysis.

​

When to use Behaviors

Use Behaviors when you need to answer questions like:

• What are users actually trying to do?
• Which assistant responses repeatedly fail an evaluator or policy?
• Which production pattern should Improve target next?
• Which traces should become regression tests?
• Did a deployment change the shape of production traffic?
• Which tool or retrieval behavior appears repeatedly before failures?
• Is an issue a one-off trace or a persistent traffic pattern?

​

Behaviors in the improvement loop

Behaviors sit between raw traces and prompt changes:

1. Monitor shows a metric shift.
2. Traces show exact requests and spans.
3. Behaviors group the repeated pattern behind those traces.
4. Improve targets the Behavior and proposes prompt candidates.
5. Evaluators and Datasets turn the issue into repeatable checks.