Coding agent insights dashboard

The dashboard at Dashboards > Coding agent insights shows your team’s coding-agent usage, estimated costs, and trace coverage. To learn how to set up tracing and a daily Loop automation that identifies recurring problems, read Improving coding agents with coding agent insights.

​

Requirements

To show data, the dashboard needs:

• Coding-agent traces sent by the bt tracing integration for Claude Code, Codex, Cursor, OpenCode, Pi, Antigravity, Grok, or Muse Code.
• A dedicated project. The dashboard counts model and tool calls from every trace in the project, including traces logged with a Braintrust SDK, so keep coding-agent traces in a project of their own.

Requires data plane v2.13.0+

​

Open the dashboard

In the project that receives your coding-agent traces, go to Dashboards > Coding agent insights. Coding agent insights is a built-in dashboard, like Cost and quality. It’s read-only, so you can’t add or edit its charts the way you can on a dashboard you create. You can star it like other dashboards. See Browse dashboards.

​

What the dashboard shows

The dashboard answers three kinds of questions about your team’s coding-agent work:

• Who is sending traces. The dashboard shows which teammates have recorded model calls matching the selected period and filters.
• How agents are used. The dashboard shows how many sessions, model calls, and tool calls your team’s agents make, and which models and tools they use.
• Where cost goes. The dashboard breaks down cost by the member or service account that authenticated each trace, as well as by permission group, model, and branch. For traces sent with an API key or service token instead of an OAuth login, it can also break cost down by key. Users, Groups, and Keys also show cost changes from the previous period.

Braintrust calculates cost from each model call’s token counts and the model’s pricing. If a span reports its own cost in metrics.estimated_cost, Braintrust uses that instead. A Usage over time chart at the top shows cost across the selected period. To plot token counts instead, click Cost beside the chart and select Tokens. Below the chart, a table summarizes usage for the period. The time range selector in the table toolbar, such as Past 7 days, sets the period for both the chart and the table. Add filter narrows both to a subset of traces.

​

Group the table

The table shows one grouping at a time. To change it, use the grouping selector in the table toolbar, which shows Users by default. Each grouping changes what a row represents and which columns appear: When a trace records both identities, it appears in both breakdowns. Users groups it by the member or service account that owns the credential. Keys groups it by the specific API key or service token used.

​

See who is sending traces

In the Users grouping, the User column header shows how many organization members are sending traces. For more detail, click the count to open the User trace coverage display, which lists who is Sending traces and who is Not sending traces. A count such as 4 / 12 means that 4 of the organization’s 12 human members sent model calls matching the selected period and filters. Service accounts are excluded from both numbers. If someone sends traces using a service account’s credentials, the dashboard attributes those traces to the service account, not that person.

​

Compare permission groups

In the Groups grouping, a member or service account that belongs to more than one permission group counts toward each group. As a result, the group rows can add up to more than the summary totals, which count each member or service account’s usage once.

​

Investigate a row

Group and branch rows open Logs, filtered to that row and the dashboard’s period and filters. For a permission group, Logs shows traces from the group’s active members. Rows for groups with more than 50 active members can’t be clicked. To open Logs in a new tab, hold Cmd or Ctrl while you click. To ask Loop about a row instead, hover over the row’s name until the Loop icon appears, then click it. Loop opens with a prompt about that user, permission group, key, model, tool, or branch.

​

When the dashboard is empty

Until the project has coding-agent traces, the dashboard shows a setup screen instead of data. It includes a command for enabling tracing in your coding agent and a link to the docs. The dashboard appears as soon as the first trace arrives.

​

Protect sensitive data

Coding-agent traces can contain source code, file paths, prompts, command output, and any credentials that appear during a session. Braintrust supports two ways to keep sensitive data out of stored traces, and you can use them together:

• Ingestion redaction (private preview) replaces detected sensitive text before storage. Model-based detection can miss values.
• Span plugins remove values on each computer before bt sends them, so those values never reach Braintrust. Span plugins require bt v0.21.0 or later.

​

Next steps

• To learn how to set up tracing and recurring problem detection for your team, read Improving coding agents with coding agent insights.
• Learn how to review and act on patterns.
• Read about built-in and custom dashboards.
• Enable tracing for each coding agent with bt trace enable.