The answer
A trace can reveal an instruction violation or a changed tool path. To determine whether behavior has shifted over time, compare similar sessions, users and intents across production windows.
Flowlines receives traces through OpenTelemetry, Langfuse, or LangSmith and analyzes them across sessions. The Issues view includes drifting as a behavior category and separates it from reliability errors.
The classes of drift worth alerting on:
Instruction drift. Policy violations, off-topic responses or scope creep become more common in comparable sessions. Tool-path drift. A formerly direct task begins looping through repeated calls. Outcome drift. Technically successful execution stops producing a useful result.
Example
A repeated tool path that still reports success
A tool journey reports successful calls but repeats the same action many times. Infrastructure health looks normal, yet the loop count and expanded path show that the agent is no longer progressing efficiently.
Flowlines links the journey summary to the use case, tools, outcome, and underlying sessions so the team can inspect the change in context.
How it works
Pattern first, evidence second
Flowlines groups related behavior into an issue and describes what happened, why it matters, and the next step. Technical detection details explain where the issue fires.
Evidence sessions remain attached to the issue, which lets an operator validate the pattern before changing the prompt, tool policy, retry ceiling, or stop condition.
Last reviewed 2026-09-04