I want to turn my proof-of-concept multi-agent help desk into a production-ready service that plugs straight into Salesforce. The current design splits the work between four specialised agents: • Classifier routes each incoming message. • Researcher pulls the relevant knowledge. • Responder drafts the reply. • Escalator monitors sentiment, flagging frustrated users so a human can jump in. n8n sits at the centre, handing tickets from one agent to the next in real time. Most of our traffic is “general information” queries, yet I need the flow to recognise when the mood sours and pass control to a live agent without delay. Once a human resolves the issue, the system must store that outcome, learn from it, and improve the next automated response. Salesforce is our only CRM, so every conversation should open or update a case, fetch the customer’s previous interactions and their stated preferences, and surface that context back to the agents. No other CRMs are involved. Deliverables I will sign off on: 1. n8n workflows covering intake, classification, research, response and escalation. 2. Bi-directional Salesforce connector that creates/updates cases and pulls interaction history and preference data. 3. Sentiment module with a configurable threshold for human handoff. 4. Feedback loop that retrains or fine-tunes the Classifier and Responder agents from resolved tickets. 5. Deployment guide and a brief video walk-through showing the system handling at least ten live test messages, including one forced escalation. Acceptance criteria: • 90 % or better accuracy on routing “general information” vs. escalate. • Handoff from Escalator to a human in under five seconds. • New knowledge incorporated and reflected in subsequent answers after one feedback cycle. If you have direct experience with n8n, multi-agent architectures, and Salesforce APIs, I’m ready to see how you would stitch this together.