Last updated: November 2026
When several agents run at once, use the Agent Graph to see what each one is doing. The graph shows every task, the sessions in each task, each turn, every tool call, every subagent, and the files the agents touch. It updates live, and new nodes appear as the agents work. The Agent Graph shows agent work the way the Commit Graph in GitKraken Desktop shows a repository’s history.
Open the Agent Graph
You can open the Agent Graph for all your agents or for a single session. It has no page of its own:
| To see | Do this |
|---|---|
| All your agents | Select the Agent Graph view on the Dashboard. To show any Tasks view as a graph, choose Graph in its View menu |
| One session | On the task page, open the menu of the session’s row in the context panel and choose Show Agent Graph. The session’s graph opens beside its transcript, and Hide Agent Graph closes it |
To learn about the Dashboard’s views and their List, Board and Graph layouts, see Arranging Your Work.
A graph view shows the tasks that its search and filters let through. A session’s graph opens fully expanded, top to bottom, and grows as the agent works.
The graph is for viewing, not for working. To work in a session, open the session from its task. The node details in the graph link directly to it. To watch several agents at once and answer them, Shift-click sessions to open them side by side. See the task page.
How to read the graph
The graph has layers. Each layer is one level further from the root:
| Layer | What it is |
|---|---|
| Task | The root, 1 for each task. See Tasks and Resources |
| Session | Each agent session in that task, with its agent and model |
| Turn | One exchange: Turn 1, Turn 2, and so on |
| Tool call | What the agent did in that turn |
| Subagent | A subagent the agent started, with its own tool calls beneath it |
| File | The files a call touched. This is the bottom layer, where you can see when 2 sessions work on the same file |
The graph shows only sessions that belong to a task, so an external session that no task owns does not appear.
Past turns start folded
In a Dashboard graph view, each session shows its latest turn and folds the earlier ones. A folded card shows a chevron and the number of nodes beneath it. To fold or unfold any card with children, such as a session, a turn or a subagent, double-click it. A session’s own graph starts with every turn unfolded.
Look out for 2 kinds of node, because the agent cannot continue until you respond: Permission needed and Question for you.
What a node’s state means
| State | What is happening |
|---|---|
| Running a tool | Working now |
| Waiting on the model | Thinking |
| Blocked | Waiting on you |
| Finished — unread | Done, and you have not looked at it yet |
| Done | Finished, and you have looked at it |
| Stopped | You interrupted it |
| Failed | It ended with an error |
| Idle | Doing nothing |
For 3 states that need your attention, the card shows a word instead of a color: New, Needs you, and Error.
Tool calls are colored by what they do
Each tool call has a color for its kind:
Read · Edit · Delete · Move · Search · Execute · Think · Fetch · Switch mode · Other
The colors are the quickest way to see what an agent is doing. A task with only Read calls is still exploring the code. Many Edit calls mean the agent has chosen an approach and is making changes.
Repeated calls of the same kind fold into one node marked Group. The node keeps the kind as its name and shows the number of calls and how many different files they touch. For example, a Read node that reads 9 calls · 4 files did not read the same file 9 times. To unfold a group, double-click it. Double-click it again to fold it. A single click only selects a node, so selecting a group never rearranges the graph.
Click any node for detail
When you select a node, its details open beside the graph. The details depend on the kind of node:
- Session: the agent, state, model, mode, context used, cost, where it is running, and its stop reason.
- Tool call: the tool, the exact input, the files and paths it touched, how long it took (some calls show no duration), and how many calls failed.
- Subagent: its type, its prompt, and its own calls.
- Task: its status (Active, On hold, Later, or Archived), its sessions, and its resources.
- File: a preview of the file. For a large file, the preview reads only its beginning.
For a node with children, the details header also has a fold button, Collapse what is under this node or Expand what is under this node. It does the same as double-clicking the card. The stats column still counts the whole graph, including folded nodes.
From the details, click Open task or Open session to go to the work, or Copy to copy any input or prompt.
The stats column
The stats column, on the right side of the graph, shows numbers about all your agents’ work. Read it from top to bottom.
On the Dashboard, the stats column hides while a panel is open.
The counters
The top of the stats column has 6 tiles, 2 across:
| Counter | What it counts |
|---|---|
| Working | Sessions in a turn, with a tool running |
| Blocked | Sessions that cannot continue without you |
| Thinking | Sessions in a turn with no tool running, waiting on the model |
| Unread | Sessions that finished a turn you have not looked at |
| Subagents | Subagents running now |
| Failed calls | Tool calls that ended with an error, across the graph shown |
Blocked and Unread are the counters that tell you to act.
Needs attention
The Needs attention list appears under the counters when something needs you. It lists everything that is waiting for you, most urgent first, tagged Needs you, Error, or New. Click a row to go to its node. In a graph with 300 nodes, this list is the reliable way to find the nodes that need you.
Throughput
The Throughput section shows a trace of your agents’ activity over time, with these readings under it:
| Reading | What it tells you |
|---|---|
| Tasks / Sessions | How many the graph shows |
| Tools | Tool calls running now |
| Files touched | How many different files the calls in the graph have touched |
| Context | How much context window your sessions have used, as a total and a percentage. Only sessions that report usage count, and the row does not appear when none do |
| Cost | What the work has cost so far, in the currency your sessions report. The row does not appear when sessions report no cost, or report it in several currencies |
Tool spectrum
The Tool spectrum section shows the calls by kind, with a bar and a count for each kind. It shows whether your agents are mostly reading, editing or running commands. Kinds with no calls do not appear, and until there are calls, the section reads No tool activity yet.
Usage windows
The Usage windows section shows your agent subscription windows, with 1 bar for each window of each agent, and the share of the window you have used. The section appears when it has something to show:
- With Check plan usage with each agent’s provider on (the default, in Settings → Agents → Usage limits), Kepler asks each agent’s provider for your usage.
- With it off, the section shows only readings that your agents report themselves, dimmed when they are out of date. The section stays hidden until an agent reports a reading.
See Settings.
These bars show only how much of each window you have used. To see whether you are on pace for a window, and when it runs out at your current rate, check a session’s usage meter. See Usage indicators.
What was folded away
At the bottom of the stats column, a note counts what Kepler left out of the graph to keep it readable:
- Tool calls grouped.
- Sessions filtered out.
- Sessions Kepler can see but not trace.
- Tasks with no session.
Finding things in a busy graph
In a Dashboard graph view, use the Dashboard’s search box and the view’s View menu. In a session’s own graph, use the search box, Options menu and depth slider in the graph’s header. The zoom buttons and the replay bar float over the bottom of the graph:
| Control | What it does |
|---|---|
| Detail depth | How many layers deep the graph goes: up to 5, and 4 by default |
| Direction | Draws the tree Left to right or Top to bottom |
| Zoom to selection | Zooms to the selected node and its tree. The While zoomed settings below it set when the view moves again |
| Group repeated tool calls | On by default. When off, every call is its own node |
| Compact stats | Shows the stats column in less space, for a smaller window |
| Sessions | On the Dashboard, under Filters. Active, the default, hides sessions that have ended |
| Only active sessions | In a session’s own graph, hides sessions that have ended. Off by default |
| Agent / Repo | In a session’s own graph, narrows it to 1 agent or repository |
| Zoom in / Zoom out / Fit to view | Moves the view. You can also drag to pan, and press Escape to clear a selection |
Each graph view keeps its own settings.
Pause and replay the graph
A live graph changes while you read it. To hold the picture still, or to look back at what just happened, use the replay bar over the bottom of the graph:
| Control | What it does |
|---|---|
| Play/pause button | From live, holds the picture on screen. While you hold a frame, plays the recording forward from it, or pauses it again |
| Previous frame / Next frame | Steps back or forward 1 frame. From live, Previous frame shows the moment just before the latest one |
| Scrubber | Moves to any point in the recording. Drag it, click it, or use the arrow keys |
| Readout | While you hold a frame, how far behind live the picture is, such as -12s. While the graph is live, how much history you can move back through |
| Live / Back to live | Reads Live while the graph follows your agents, and Back to live while you hold a frame. Click Back to live to return to live |
To see the time of the last update, hover over Live.
Your agents keep working while you hold a frame, so the readout shows the picture falling further behind. Kepler records only while an agent is working, so while every agent is idle, the replay bar’s controls are unavailable.
What the graph does not show
The graph does not show everything about your agents’ work:
- No token counts per turn. The graph shows how full the context window is, and the cost, instead.
- No durations for some calls. See Click any node for detail.
- No cost or time per subagent.
- No activity for external and terminal sessions. The graph shows each one as a single node with a badge, instead of showing it as idle.
- No history after Kepler restarts. After a restart, the nodes beneath a session are empty, though the session node remains.
When nothing is running, the graph reads Nothing running yet, with the hint Start an agent session and its activity appears here in real time. When sessions exist but your filters exclude all of them, the hint changes to No sessions match the current filters.
Send feedback
The Agent Graph is new, and it will change. To tell the Kepler team whether it helps you, click Feedback in the top bar to open Send feedback, then follow these steps:
- Pick a Type: Feature request, Bug report, or General feedback.
- Write a Message.
- Send it.
You can also post feedback on the public issue board at github.com/gitkraken/gk-ade/issues.
