A tenant-scoped (per-organization) AI oversight layer that acts as a "manager" for worker agents. Instead of treating an agent's last message as the end of the job, this Manager Agent reviews the interaction to ensure quality, security, and completeness.
- Oversight: Catch mistakes, dropped action items, and hallucinations.
- Actionability: Trigger internal follow-ups (Slack/Email/Jira) when needed.
- Security: Prevent prompt injection and ensure data isolation between orgs.
- Input: Full conversation transcript (User + Agent) + Tool Execution Logs.
- Metadata: Org ID, Customer Tier, Active Tools.
- Role: Evaluator.
- Prompt Strategy: "You are a QA Manager. Review this transcript. Did the agent solve the user's intent? Did it use tools correctly? Check for security risks."
- Output Structure:
{ "verdict": "PASS" | "FOLLOW_UP" | "ESCALATE", "reason": "Agent failed to ask for order ID before checking status.", "security_risk": "None", "suggested_actions": ["email_support", "log_ticket"] }
- Internal Comms: Send alerts to internal Slack channels or Support Email.
- Sanitized Web Search: Verify facts if the agent made claims (e.g., "Is our API actually down?").
- CRM/Ticket Update: Draft a Jira ticket or update HubSpot if the interaction requires human intervention.
- Tenant Isolation: Ensure Manager A only sees Org A's data.
- Prompt Injection Guardrails: The Manager must detect if the user tried to trick the worker agent (e.g., "Ignore previous instructions").
- Playbooks: Different evaluation rules for Support vs. Sales agents.
- Native Tooling: Can easily connect to Slack, Email, and Web.
- State Awareness: Can maintain "Org Context" over time (long-term memory).
- Security: Self-hosted / Private deployment ensures customer data doesn't leak.