What You’ll Accomplish
This guide walks you through the core Vijil workflow:
- Register your agent in the Agent Registry
- Register your target environment with personas and policies
- Create a custom harness with generated test cases
- Evaluate your agent
- Review the results and Trust Score
- Defend it with Dome guardrails
- Observe agent security and safety in production
By the end, you’ll have a Trust Score for your agent and a trusted agent ready for production.
Step 1: Register Your Agent
Navigate to Agents in the sidebar to open the Agent Registry.
Click + Register Agent to open the registration modal.
Vijil supports three access levels, each enabling progressively deeper evaluation:
| Access Level | What You Provide | What Vijil Can Test |
|---|
| Black Box | Description, API endpoint, credentials | Input/output behavior only |
| Grey Box | Model config, MCP config, A2A config | Vulnerabilities traced to agent composition |
| White Box | Full configuration and source code | SAST and DAST analysis for thorough audit |
- Enter a Name for your agent
- Add your Agent Description
- Set Status to Draft for evaluation
- Provide the Agent URL (where it’s running)
- Add your Access Key and set the Rate Limit
- Click Register Agent
If your agent delegates to sub-agents or uses MCP tools, expand Grey Box to register those connections. This enables Vijil to test the full execution graph.
Step 2: Create a Custom Harness
While the Trust Score harness provides a comprehensive baseline evaluation, custom harnesses let you test your agent against specific personas and organizational policies.
Navigate to Harnesses and click + Create Harness to open the harness wizard.
The wizard walks you through four steps:
- Basic Info — Name and describe your harness
- Select Agent — Choose which agent this harness evaluates
- Select Personas — Choose user profiles that will interact with your agent
- Select Policies — Choose organizational rules your agent must follow
Choose Personas
Personas define who interacts with your agent during evaluation. Vijil generates realistic test cases from each persona’s perspective.
| Persona Type | What It Tests |
|---|
| Regular users | Normal usage patterns, edge cases |
| Security researchers | Adversarial prompts, jailbreak attempts |
| Domain experts | Technical accuracy, specialized knowledge |
Choose Policies
Policies define what rules your agent must follow. These can be:
- Operational policies — Business rules, response guidelines
- Compliance frameworks — GDPR, HIPAA, SOC 2 requirements
- Safety policies — Content restrictions, escalation procedures
Once configured, click Create to save your harness. Vijil generates test cases based on your personas and policies:
Review Your Harness
After creation, the harness detail page shows your configuration:
Set status to Active when ready to use in evaluations.
Step 3: Run an Evaluation
Navigate to Evaluations to open Diamond Evaluations.
The evaluation page presents two panels:
Select Agent — Choose which registered agent to evaluate.
Select Harness — Choose what to test. The Trust Score harness evaluates three dimensions:
| Dimension | What It Measures |
|---|
| Reliability | Correctness, consistency, robustness |
| Security | Confidentiality, integrity, availability |
| Safety | Containment, compliance, transparency |
Monitor Progress
While the evaluation runs, the status updates in real-time in the Evaluation Results table. Depending on the rate limit, evaluations can take 5-30 minutes.
Step 4: Review Results
When the evaluation completes, click the view icon in the Actions column to open the Trust Report.
The report provides:
- Trust Score — A 0-1 score (threshold: 0.70 for passing)
- Pass/Fail Status — Clear deployment recommendation
- Per-Harness Breakdown — Scores for each evaluated dimension
- Agent Specification — Configuration used during evaluation
- Deployment Recommendation — Action items based on findings
Interpreting the Score
| Score | Interpretation | Next Step |
|---|
| ≥ 0.70 | Passed | Deploy with standard monitoring |
| 0.50 - 0.69 | Marginal | Deploy with Dome guardrails |
| < 0.50 | Failed | Remediate before production |
The Trust Score quantifies known risks based on the harness you ran. It doesn’t guarantee absence of all vulnerabilities—only that your agent performed acceptably against the tested scenarios.
- Security Guard — Detects prompt injection, jailbreak attempts
- Moderation Guard — Blocks toxic or inappropriate content
- Privacy Guard — Redacts PII from inputs
Output Guards — Filter responses before they reach users. Common guards:
- Privacy Guard — Prevents PII leakage
- Moderation Guard — Catches harmful outputs
- Click + Add in either Input Guards or Output Guards
- Select a guard type from the dropdown
- Configure execution settings:
- Early Exit — Stop pipeline on first detection (recommended for security)
- Serial/Parallel — Execute guards sequentially or concurrently
Test Your Configuration
Use the Execution Flow panel to test how inputs flow through your guardrail pipeline:
- Enter a test prompt in the input field (e.g., a message containing PII)
- Click Send to see how each guard processes it
- Watch the flow diagram show which detectors activate
- Verify guards trigger as expected for your test cases
The visual flow shows each guard and its underlying detectors—like prompt-injection-mbert for security or privacy-presidio for PII detection.
When satisfied, click Save Configuration.
Step 6: Deploy and Monitor
With Dome configured, integrate it into your application using the SDK:
from vijil import Dome
dome = Dome(
api_key="your-vijil-api-key",
agent_id="your-agent-id"
)
# Guard inputs before sending to your agent
safe_input = dome.guard_input(user_message)
# Guard outputs before returning to users
safe_output = dome.guard_output(agent_response)
Monitor in Production
Return to the Guardrails dashboard to monitor your protected agents. Click View Dashboard to see:
- Real-time detection events
- Guard trigger rates
- Blocked vs. allowed requests
Next Steps
