> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vijil.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Evaluate Agents

> Run behavioral and Red Team evaluations against Agents with the CLI.

Diamond supports two evaluation types. A **behavioral evaluation** uses [Harnesses](/concepts/evaluation-components/harness) to send [Probes](/concepts/evaluation-components/probe) and produce a [Trust Score](/concepts/trust-score/introduction). A **Red Team evaluation** runs adaptive attack waves against the Agent and produces findings and a report. Both use an evaluation ID.

## Run an Evaluation

Use `vijil evaluate <agent>` with an Agent ID or alias. The default type is `behavioral`. Choose either the standard Harness set or a custom Harness:

```bash theme={null}
vijil evaluate "$AGENT_ID" --baseline
vijil evaluate "$AGENT_ID" --bespoke "$HARNESS_ID"
```

Set `--type redteam` to start a Red Team evaluation. Pass an optional JSON object with `--redteam` to override wave settings:

```bash theme={null}
vijil evaluate "$AGENT_ID" --type redteam
vijil evaluate "$AGENT_ID" --type redteam --redteam '{"max_waves": 3}'
```

The command waits for a terminal status and prints the result by default. Add `--no-wait` to create a cloud evaluation and return its ID immediately:

```bash theme={null}
vijil evaluate "$AGENT_ID" --type redteam --no-wait
```

| Option | Applies to | Description |
| - | - | - |
| `--type <type>` | Both | `behavioral` or `redteam`; defaults to `behavioral` |
| `--baseline` | Behavioral | Use the standard trust Harness set |
| `--bespoke <harness-id>` | Behavioral | Use a custom Harness |
| `--sample-size <number>` | Behavioral | Set the number of Probes per Harness |
| `--redteam '<json>'` | Red Team | Override Red Team settings with a JSON object |
| `--no-wait` | Cloud evaluations | Return after creating the evaluation |
| `--local` | Behavioral | Run against a locally registered Agent; see [local evaluations](/developer-guide/evaluate/local-agent) |
| `--harness-name <name>` | Local behavioral | Select `safety`, `security`, or `reliability`; defaults to `safety` |

Red Team does not use `--baseline`, `--bespoke`, `--sample-size`, `--local`, or `--harness-name`.

## Inspect Results

List evaluations for an Agent, or show a behavioral evaluation by its ID:

```bash theme={null}
vijil evaluations list --agent "$AGENT_ID"
vijil evaluations show "$EVALUATION_ID"
vijil scores show "$AGENT_ID"
vijil reports download "$EVALUATION_ID"
```

`vijil scores show` returns the Agent's latest Trust Score. For Red Team progress and reports, use the evaluation-scoped commands:

```bash theme={null}
vijil evaluations redteam-status "$EVALUATION_ID"
vijil evaluations redteam-report "$EVALUATION_ID" --format md
```

See the [Red Team CLI guide](/developer-guide/cli/red-team) for seeds, attacks, judgments, reflections, and report formats. To request JSON output from a CLI command, use the global option before the command, for example `vijil --output json evaluations list`.

## Manage Behavioral Harnesses

Harnesses apply to behavioral evaluations. The current CLI provides these resource commands:

| Command | Description |
| - | - |
| `vijil harnesses list` | List available Harnesses |
| `vijil harnesses show <harness-id>` | Show Harness details |
| `vijil harnesses create` | Create a custom Harness for an Agent |
| `vijil harnesses update <harness-id>` | Rename a Harness |

Create a custom Harness, then pass its ID to `--bespoke`:

```bash theme={null}
vijil harnesses create --name "Customer Support" --agent-id "$AGENT_ID"
vijil evaluate "$AGENT_ID" --bespoke "$HARNESS_ID"
```

`vijil harnesses create` also accepts `--description`, `--system-prompt`, `--system-prompt-file`, `--persona-ids`, and `--policy-ids`. Separate Persona or Policy IDs with commas.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.