Dashboard Guide

Agents

Subagent usage by type — how often each one runs, what it costs, and what it is connected to.

What the Agents pages show

Agents in the sidebar covers the subagents Claude Code launches with its Agent tool: Explore, general-purpose, a plugin's engineering:reviewer, your own project agents, and so on. It answers:

  • Which subagents do my sessions actually use, and how often?
  • What does each one cost per call, and how long does it take?
  • What launches a given subagent, and what does it use once it's running?

Everything is grouped by agent type, the name Claude Code reports for the subagent. An invocation with no type is counted as general-purpose, which is what Claude Code runs when none is given.


Agents list

Agent Calls Over Time

A daily count of finished subagent runs. It counts subagent_stop events, so a subagent that was still running when its session's events were sent isn't counted until a later flush delivers its stop.

Top Agents

The top 50 agent types in the selected range, by number of calls. Click a type to open its detail page.

ColumnWhat it shows
Agent TypeThe subagent's type. Links to the detail page
OriginWhere the agent is defined — see below
CallsFinished runs in the period
Avg CostEstimated cost per call
Tokens (in/out)Input and output tokens across all calls
Latency p50 / p99Median and tail time from launch to result
Avg Tools/CallTool calls per run
Avg TurnsModel turns per run
TrendChange in calls against the preceding period of the same length

The chips above the table filter by origin. Only origins that appear in the range are shown.

OriginMeaning
PluginDefined by a Claude Code plugin. The type is namespaced, like plugin:agent
ProjectDefined in the project's .claude/agents/
UserDefined in your own ~/.claude/agents/
Built-inShips with Claude Code, such as Explore, Plan or general-purpose
Ad-hocNone of the above could be confirmed

Project and User need the plugin to see your filesystem, so it decides them when the subagent starts. If that classification wasn't captured, for example from an older plugin, Trenchcoat can only tell plugin and built-in agents apart from the name, and a project or user agent shows as Ad-hoc.

Latency needs plugin 1.2.0 or newer. If none of the agents in range has a latency sample, the page says so above the table, and the latency columns read --. That means "not measured", not zero.


Agent detail

The detail page covers one agent type over the selected range.

Summary cards

Invocations, Total Cost, Avg Cost/Call, Tokens (in/out) and Median Latency. The latency card only reports a value when latency was captured.

Relationships

A picture of everything this agent type is connected to. The type sits in the middle.

  • Left: Spawned by. What launched it. Either the main conversation, or another agent type that ran it as a subagent of its own.
  • Right, in labelled groups:
    • Spawns: agent types it launched itself
    • Skills, MCP servers and Tools it called while running
    • Models it ran on

The number beside each item is how many times that relationship happened: invocations for spawns and models, calls for skills, MCP servers and tools. A thicker line means more. When a type launches another run of its own type, the count appears under its name in the middle ("spawns itself 5×") and as a row marked "(itself)" in both lists.

Reading it:

  • Hover or click an item to highlight its line and see the count on it. Clicking opens its details in the panel to the right: sessions, total and average run time, result size for calls, and when it was last seen.
  • Agent types, skills and tools link to their own pages, so you can follow a chain: from general-purpose to the Explore runs it launches, on to what Explore uses, and into any skill or tool it calls (see Skills and Tools). MCP servers don't link: this picture groups MCP calls by server, and a server has no page of its own.
  • The chips above the picture show and hide whole groups.
  • Long groups fold into a "+N more" item. Click it, or Draw all in the panel, to show every item.
  • Nothing selected, the panel lists every relationship as text with its share of the group, plus the most recent sessions this type ran in. Each session links to its session page. On a phone, the panel sits below the picture and is the easier way to read it.

A sentence above the picture sums it up, for example: "Spawned 131 times by 4 callers, most often the main conversation (120). Spawns 4 agent types, calls 14 tools, 3 skills and 2 MCP servers and runs on 2 models."

Main conversation also includes launches whose parent wasn't recorded. When a subagent launches another one in the background, Claude Code doesn't always report which agent did it. Those runs are counted under the main conversation, the same as in a session's call graph.

What it leaves out, deliberately:

  • Cost. The summary cards already show this type's cost. The picture shows counts and durations, so the page never shows two cost figures for the same runs.
  • Errors. Tool failures aren't yet captured reliably enough to count, so the picture shows no error counts rather than a misleading zero.

Charts

Cost per day, Invocations per day and, when latency was captured, Latency (p50) per day.


Filtering

  • Date range: the picker at the top right. Every card, chart, table and the relationship picture use it.
  • Machine: the global machine filter. Each API key maps to one machine. Following a link from one agent type to another keeps both filters.

On this page