Pipelines

A pipeline is a group of agents orchestrated to complete a task.

createPipeline('support-reply') starts one session of that task. The calls in the session share one cost, one duration, and one failure count. Run the task again and you get another session. The log page lists the pipeline once, then compares its sessions.

The agents inside support-reply are track and answer. Their calls stay in the session. See Integrate for the request fields.

On the log page

Screen	Address
Pipelines	/
One pipeline	/sessions?pipeline=support-reply
One session	/session?pipeline=support-reply&session=pipeline:support-reply-execution:…

Pipelines is one row per pipeline name. Sessions is how many times the task ran in the current window. Agents is how many agents appeared in it. That count opens a side panel of those agents.

Avg cost is the mean cost of one session. P50 cost, P75 cost, P90 cost, and P95 cost are percentiles of those session costs. Avg latency is the mean wall-clock time from the first call start to the last call end. Avg model time is the mean of each session’s summed call durations. Overlapping calls both count in model time.

Avg latency, Avg model time, and the four cost percentiles start hidden. Turn them on from Columns.

One pipeline lists sessions. Latency is that session’s wall-clock span. Model time is the sum of its call durations. Those two columns start hidden. Stats above the table include input, output, and cached tokens, plus the same cost percentiles for the sessions on screen.

One session lists that session’s model calls, including every agent in it. Latency on a call is that call’s duration. A row opens the messages panel. Functions and Server open the function-call panel. A function result is joined back to the original call by call id, so the panel can show the description and schema.

Filters are shared with Agents. They are listed on The log page.