---
title: "Judgment Agent"
description: "Ask questions about your agent's data and get rich answers."
sidebar:
  label: "Chat"
seo:
  title: "Judgment Agent | Judgment Docs"
  description: "Ask questions about your agent's data and get rich answers."
---

The Judgment Agent reads your traces, judge labels, Issues, datasets, and test runs. You ask in plain words and it writes the queries. The same agent runs [Scheduled Tasks](/documentation/agents/scheduled-tasks) and answers in Slack as the [Slack Agent](/documentation/agents/slack-agent) when someone mentions @Judgment, and it reads [Agent Knowledge](/documentation/agents/agent-knowledge) wherever it runs.

## How it answers questions at scale

The agent queries your project with Judgment's virtual SQL (vsql), the same read-only tables you can [query yourself](/documentation/agents/query-with-sql). When your judges have already labeled traces, it finds the sessions your sentiment judge labeled Negative and reads what the other judges said about those same traces, instead of re-reading every trace. Because labels are computed ahead of time, answers stay fast as volume grows. [Labels as indexes](/documentation/agents/labels-as-indexes) explains the idea.

## Where to find it

- On your project's home page, type in the box that says "Ask Judgment about lunch-order-agent. @ for context, / for skills".
- Select **Ask Judgment** on a judge, a label inside a trace, or a selection of traces or Issues to start a chat with that context attached.
- Open past conversations under **Chats** in the sidebar.

The chat panel can **Dock left**, **Dock right**, **Float**, or go **Full screen**.

{/* SCREENSHOT: Judgment Agent docked right beside a trace in the lunch-order-agent project, with an @-mentioned judge in the composer, dark mode */}

## Give it context

The page you're on comes along with your message. Inside a trace, the trace and the span you've focused are attached automatically.

Type `@` to add more. The picker has Traces, Spans, Judges, Labels, Automations, Views, Datasets, and Knowledge. Mentions stay attached for the rest of the conversation.

You can also attach files with **Attach files**, for example a CSV export from your support tool.

## Skills with /

Type `/` to run a skill. Skills are files you keep in the `skills/` folder of [Agent Knowledge](/documentation/agents/agent-knowledge). A file named `skills/weekly-lunch-review` shows up as `/weekly-lunch-review`. Use them for questions your team asks over and over, so everyone gets the same steps and the same output format.

## Fast or Deep Research

Each chat runs in one of two modes:

| Mode | Use it for |
| --- | --- |
| **Fast** | Quick lookups and summaries: "How many orders failed yesterday?" |
| **Deep Research** (default) | Questions that need several queries and some judgment: "Why were users frustrated last week, and is it new?" |

Both modes use the same tools and data. Deep Research takes longer before it answers.

Pick the mode before you send the first message. After that, the mode and the write setting below are locked for the chat. To change either one, start a new chat.

## Approving writes

Reading is always allowed. For anything that changes data, you choose:

| Setting | What happens |
| --- | --- |
| **Ask for writes** (default) | Before a tool creates, updates, or deletes something, the agent shows an approval card. Select **Allow** (the check mark) or **Deny** (the X). When several tools are waiting, **Allow all** approves the rest of them. |
| **Auto-allow writes** | Write tools run without asking. |

Your role still applies either way. A Viewer can't create a judge through the agent any more than through the app. See [Members and roles](/documentation/administration/members-and-roles).

## What it can do

| Area | Examples |
| --- | --- |
| Traces and sessions | Find slow or failing requests, compare agent versions, summarize one trace span by span |
| Judges and labels | Draft a new judge from a description, add an action like "post to Slack when this judge returns Failed silently", check how often a label was applied, explain why a trace got a label |
| Issues | Root-cause an Issue, rank open Issues by impact |
| Datasets and tests | Create a dataset, read test run results |
| Agent Knowledge | Read your team's files and keep its own notes. It picks files by their names and paths. |

A few things it can't do yet: add or remove examples in an existing dataset, or create a Scheduled Task. Do those in the app.

## Questions to try

These are written for the [lunch-order agent](/documentation/start/example-agent).

```text
How often were users frustrated with the lunch-order agent last week?
Group it by cuisine and order size, and tell me what caused it.
```

```text
Which tool calls failed most in production this week? For the top one,
show me three traces and say whether the agent recovered.
```

```text
Create a judge that flags traces where get_menu fails and the agent drops
that restaurant without telling the user. Show me the rubric before you save it.
```

## Find and share chats

Your chats are listed under **Chats**. Open **Chat display options**, then **Show chats from**, to switch between **Personal** (yours) and **Team** (chats shared in the project).

From **Chat actions** on a conversation:

| Action | What it does |
| --- | --- |
| **Copy session link** | Copies a link. Only people with access to the project can open it. |
| **Share with team** | Makes your chat visible to teammates in the project. Shows **Make private** once shared. |
| **Copy as prompt** | Copies the conversation as a prompt you can paste into Codex, Claude Code, or Cursor to keep working on a fix. |

Teammates can read a shared chat but can't add to it.

## Outside the chat

The same agent also runs Issues detection, with the same tools everywhere. In Slack, it doesn't stop to ask before writes.

## What it costs

The agent's work uses Judgment Units (JU). It shows up as **Judgment Agent** in the usage breakdown, under the **Evaluations/platform agent** workload on the billing page. See [Billing and plans](/documentation/administration/billing-and-plans).