Why profiles matter
Profiles give the evaluation engine context. A support bot and a code-review bot should be judged differently. The profile’s role and description inform dimension scoring.Create a profile
- Go to Agents in the sidebar
- Click Create profile
- Fill in:
- Name — display name (e.g. “Acme Support Bot”)
- Role — what the agent does (e.g. “Tier-1 customer support”)
- Description — policies, boundaries, escalation rules
- Visibility — private or public (see Public profiles)
Link to your code
Profiles connect to runtime via API keys:- Create or edit an API key in Settings
- Link the key to the profile
- Send decisions with a matching
agent_idin the SDK
Profile dashboard
Each profile page shows:- Decision and evaluation stats
- Dimension breakdowns
- Trend over time
- Link to public profile (if enabled)
Best practices
Write descriptions like a brief for a human reviewer
Write descriptions like a brief for a human reviewer
Include escalation rules, forbidden actions, and domain boundaries. The judge model uses this text.
One profile per distinct agent product
One profile per distinct agent product
Don’t merge unrelated agents — scores become meaningless.
Update descriptions when policies change
Update descriptions when policies change
Re-evaluations use current profile text for new decisions; keep it in sync with production prompts.