Overview

Understand recurring customer and agent patterns from logs, traces, sample conversations, and production evidence

Behaviors is Adaline’s pattern layer. It turns production logs into recurring user intents, assistant habits, tool patterns, issue diagnoses, and improvement targets.

Use Behaviors when a customer report, Monitor change, log search, evaluator failure, or support escalation looks bigger than one request.

The catalog is the first triage surface. It shows the current analysis snapshot, active filters, issue-focused saved views, related objects, and one row per repeated pattern. Use it to decide which patterns deserve evidence review, not to make the final decision from the label alone.

What a Behavior represents

A Behavior is a recurring pattern detected from project evidence. Depending on what you log, it can represent:

• A repeated user goal or support intent.
• A repeated assistant response pattern.
• A tool, retrieval, or function-call pattern.
• A failure mode, drift, or correction opportunity.
• A coding-agent pattern such as editing, testing, recovering, looping, or skipping verification.

Treat the label as a starting point. The evidence on the detail page is what turns a label into a decision.

The workflow

1. Send useful log traces and spans with proper names, status, input/output content, metadata, and other safe context.
2. Let Adaline detect recurring patterns across traces, spans, model outputs, tool calls, outcomes, and metadata.
3. Review Behavior patterns in the catalog by issue signal, role, volume, member count, and evidence.
4. Open representative evidence to inspect source logs, spans, sample conversations, and detail analytics.
5. Choose an action: watch, save a view, add coverage, start Improve, tune evaluators, debug tools, fix backend logic, or roll back.

What to read first

Start with the saved view and filters so you know whether you are looking at all patterns, issue-only patterns, or a volume-sorted view. Then read the role, issue tag, short summary, and member count on each row. A high-volume non-issue Behavior can be a healthy workflow worth preserving; a low-volume issue Behavior can still matter if it affects a critical release path.

When you open a Behavior, use the detail page to confirm the pattern with sample conversations, source traces, timing, and trend evidence. For coding-agent projects, open the session or trajectory view when the question is about the journey: planning, editing, tool use, verification, recovery, or final handoff.

Key objects

Object	Meaning
Trace	One request or run with model, tool, retrieval, backend, cost, latency, token, tag, and attribute evidence.
Span	A meaningful step inside a trace, such as a model call, tool call, retrieval step, backend call, or verification step.
Behavior	A repeated pattern across traces, spans, outputs, tool calls, outcomes, or metadata.
Sample conversation	A representative example that contributed to a Behavior and can be opened in Traces for source evidence.
Triage finding	A diagnosis attached to a Behavior when Adaline has enough evidence to explain the issue and suggest a fix.

For coding agents and other multi-step workflows, Adaline can also use trajectory-style evidence to explain the journey behind a Behavior. Treat that as an extra layer for the Coding-agent Behaviors and Trajectories workflows, not as the default requirement for all Behavior analysis.

How to prioritize

Do not treat every Behavior as a bug. Start with the role to see whether the pattern is driven by the user, assistant, tool, or agent workflow. Then look for the issue tag, which means Adaline has flagged the Behavior as likely problematic or worth review. Use member count to understand how much evidence is behind it, and compare newer Behavior snapshots to older ones to see whether similar patterns keep reappearing. A repeated pattern across snapshots may be worth monitoring or preserving even when it is not an issue.

The best Improve targets are usually specific issue rows with enough representative evidence. Healthy repeated patterns are better for monitoring and regression coverage.

Where to start

Goal	Page
Send logs that produce useful Behaviors	Logs to Behaviors
Read the catalog and detail views	Understanding Behaviors
Understand coding-agent patterns	Coding-agent Behaviors
Review multi-step journey evidence	Trajectories

Logs to Behaviors

Instrument traces, spans, tools, agent identity, and safe metadata.

Understanding Behaviors

Read issue views, volume, sample conversations, heatmaps, timelines, and actions.

Coding-agent Behaviors

Understand task trajectories, phases, tech stack tags, and failure modes.

Trajectories

Connect coding-agent and multi-step patterns back to journey evidence.