← All work

Multi-agent assistant with human approval

A multi-agent system where an orchestrator routes requests to sales, support and research agents, and people approve sensitive actions before they reach a customer.

AgentsOrchestrator + 3 specialists
ControlHuman-in-the-loop
SecurityPrompt-injection defence

The challenge

The business wanted AI to handle sales, support and research requests, but with full control over what reaches customers and protection against misuse.

What we built

  • An orchestrator agent classifies intent and routes to Sales (lead qualification), Support (RAG problem solving) and Research (information synthesis) agents
  • A FAISS-based RAG system for grounded retrieval
  • Input validation, spam filtering, prompt-injection detection and rate limiting
  • A human-in-the-loop dashboard where operators approve, reject or override agent actions
  • Monitoring through LangSmith traces

How it works

  1. User via web or API
  2. Validation and spam filter
  3. Orchestrator: intent
  4. Sales · Support · Research agents
  5. Operator approval

Results

  • Automation with a human safety net for every customer-facing action
  • Traceable decisions through LangSmith
  • Protected against prompt injection and abuse