Skip to content

The open-source agent harness - the runtime layer that turns an LLM into a working agent.

2.8 KB
Raw Blame
---
title: 'Sessions'
sidebarTitle: 'Sessions'
description: 'Review past agent runs: turns, tool calls, subagents, tokens, and timing.'
---

**Sessions** records every conversation with an agent. Open it from the sidebar to see what an agent did on a run, how long each turn took, which tools and [subagents](/key-features/subagents) it used, and where something went wrong.

## The sessions list

Every session is listed newest first. Each row shows the opening message, when it ran, how many turns it took, and its total duration. Use the **agent** and **time window** filters at the top (the last 30 days by default) to narrow the list to recent activity or to a single agent.

Runs started by a [schedule](/schedules) appear here too, marked **Scheduled run** and showing the agent that ran.

<Frame caption="Agent Sessions: recent runs with their turn count and duration, filterable by agent and time window.">
  ![The Agent Sessions list showing recent runs, each with its opening message, time, turn count, and duration, with the Agents filter open on an All and a per-agent option and an empty detail pane reading Select a session to view details](/images/sessions-list.png)
</Frame>

## Session detail

Click a session to open it. The detail view has three parts.

### Summary metrics

A row of totals for the whole session: **Turns**, **Duration**, **Tokens**, **Context**, **Tool calls**, **Sub-agents**, and **Errors**. It summarizes the run at a glance.

### Event timeline

A timeline across the run, colored by event type, such as **System**, **User**, **Model**, **Tool call**, and **Sub-agent**. It shows how the run's time was spent, so a slow model call, a long-running tool, or parallel subagents are easy to spot.

### Turn-by-turn transcript

Below the timeline, each turn is listed with its own token count and duration, the user message, and an **Agent steps** entry you can expand to see the tool calls the agent made and how they resolved.

Together they make it easy to pinpoint where a run failed, whether a turn was cancelled, an [MCP server](/mcp-servers) errored, or a tool returned nothing useful.

<Frame caption="A session opened for inspection: summary metrics, the event timeline, and the turn-by-turn transcript.">
  ![A session detail view with summary metrics (turns, duration, tokens, context, tool calls, subagents, errors), an event-type timeline, and the turn-by-turn transcript with expandable agent steps](/images/sessions-detail.png)
</Frame>

## Resuming a run

Each session carries a copyable **session ID**, and the detail header has a button to **resume** it, which reopens the conversation so you can continue it.

<Note>
  In hosted mode with [login](/authentication/overview) enabled, session history is owner-scoped: each user sees only their own sessions. In local mode there is a single shared identity, so all sessions belong to that one user.
</Note>