Architecture
One agent platform. Seven use cases. Humans in control.
The reason we recommend one pilot rather than seven projects: they share a backbone. Build the agent + MCP + connector layer once, and each use case becomes a configuration on top — on the same Vercel + Supabase + Claude stack we run in production today.
People — scholars, program staff, coordinators, comms, HR
Web console on Vercel + Ask-Andy chat
▼
Agent & orchestration layer (Claude)
extract · summarize · dedupe · monitor · draft · schedule
▼
Concur
Expense posting (MCP)
Salesforce
Contact CRM
Journal databases
Library research
Marketo / events
RSVP + comms
Andy (Intranet)
HR & Comms docs (RAG)
Web / social
QC monitoring
▼
Supabase — Postgres · Auth (roles) · Row-Level Security · audit trail
systems of record + agent memory
Agent & orchestration layer
- Claude-powered agents: extract, summarize, dedupe, monitor, draft
- Scheduled background agents (KMS updater, QC scan, RSVP chaser)
- Task planning + tool use, one shared orchestration core across all 7 cases
MCP servers & connectors
- Custom MCP server for Concur (and future systems)
- Secure connectors: Salesforce, journals, Marketo, Andy, web
- Credentials vaulted — agents act with scoped, revocable access
Human-in-the-loop
- Agents propose; people approve — every post, merge, and send
- Confidence scores + flags surface what needs a second look
- Nothing published or written to systems of record without sign-off
Security first
Scoped credentials, RLS, and audit logs. A think tank's data governance is non-negotiable.
Your data stays yours
Runs in your cloud/tenancy; models don't train on your content.
Incremental
Ship one pilot, measure, then light up the next use case on the same platform.