From Canvas to Production: How Agent Deployment Works
You've built the flow in Agent Canvas. Here's what happens when you press deploy — phone number assignment, traffic routing, monitoring, and iteration.
The gap between a working agent in the builder and a production agent handling real calls is where many deployments stall. The canvas looks good, test conversations work, but the operational questions pile up: how do I assign a phone number? Can I route only some calls to the agent? How do I monitor what's happening? What if I need to change something after go-live?
Deployment options
- Phone number — provision a new number or port an existing one. Set the agent as the handler for all inbound calls on that number.
- Web embed — deploy the agent on your website via a widget. Callers click to talk from their browser using WebRTC — no phone network needed, lowest latency.
- Call forwarding — keep your existing phone number and forward to the agent number when you don't answer (overflow) or outside business hours.
- SIP integration — connect to your existing PBX or contact center via SIP trunk for enterprise environments.
Gradual rollout
Don't send 100% of traffic on day one. Start with after-hours only, or a specific call type, or 10% of inbound via routing rules. Monitor the analytics dashboard for resolution rates, escalation patterns, and caller satisfaction. Fix issues on the canvas in real time — changes deploy instantly without downtime. Expand traffic as metrics prove the agent is performing.
Iteration never stops
A deployed agent is not a finished agent. The analytics surface knowledge gaps, conversation dead ends, and action failures every day. Each one is a canvas fix — add a new knowledge base entry, add a branch for an unhandled intent, adjust a guardrail threshold. The best deployments treat the canvas as a living system that improves weekly, driven by real production data.
Ready to build?
See how Mazed's multimodal AI agents work for your use case.