ENTERPRISE · INTERNAL + USER-FACING AGENTS · 2026
Internal triage and user self-service, engineered in parallel.
Confidential · national membership organization
THE BRIEF
A client needed two LLM agents: one for internal ops triage, one for user self-service. Same agent loop and tool registry; different trust boundaries. The ops agent runs behind corporate SSO. The user agent can only access the signed-in user's data. Every write action goes through draft-preview-confirm. A full adversarial eval suite before the user agent touched a real session.
STACK
Two-agent architectureDraft-preview-confirm flowsAPI-layer authorizationAdversarial eval harness
MEASURED OUTCOMES
Agents delivered
·
2
Adversarial eval cases
·
40+
Time from spec to production
·
4 weeks
Cost per conversation turn
·
$0.02