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

# Agent Configuration

Configure mode, model, tone, prompts, guardrails, memory, tests, and improvements.

# Agent Configuration

The AI agent is configured per project. It uses project knowledge, conversation context, customer memory, guardrails, and model settings to suggest, draft, or send replies.

## Modes

| Mode | Best for |
| --- | --- |
| Assist | Teams that want suggestions but no automatic drafts. |
| Copilot | Teams that want AI drafts reviewed by humans. |
| Autopilot | Teams with strong knowledge coverage and clear escalation rules. |

## Configuration Areas

| Tab | Purpose |
| --- | --- |
| Configuration | Mode, model, tone, prompt, temperature, and response settings. |
| Test | Run controlled questions against the current agent configuration. |
| Knowledge | Inspect the articles, crawls, docs sites, and vector sources available to the agent. |
| Guardrails | Require citations, redact PII, limit answer length, and define escalation topics. |
| Improvements | Track missing knowledge and suggested changes. |
| Memory | Project-level memory, history, pending changes, and rollback. |

## Guardrails

Guardrails are evaluated before an answer becomes customer-facing. Common guardrails include:

- require citations
- redact PII
- disallow unsafe or unsupported topics
- escalate certain topics
- cap answer length
- require human approval for actions

## BYOM

RogerIQ supports bring-your-own-model flows through model provider configuration. Store model credentials only in server-side settings, never in browser code.

## Testing

Run test questions before switching to Autopilot. Use examples from real tickets and include edge cases:

- missing docs
- angry customers
- refund requests
- ambiguous account identity
- policy exceptions
- security-sensitive questions

Autopilot should be earned by test results, not enabled because the demo looked good once.