---
title: "Deploy Multi-Model Workflows"
description: "Step-by-step guide to deploy multi-model workflows with Norcaster: providers, deployments, workflow graph, routing, runtime policies, tools, traces, and evals."
canonical: https://www.norcaster.com/docs/multi-model-agents
---
<!-- Generated at build time from the rendered page by portal/scripts/generate-llms.mjs. The page is the source of truth; edit it, not this file. -->

Multi-model workflows

# Deploy multi-model workflows with Norcaster

This guide shows how to deploy an agent-style workflow where different supported steps can use different models, providers, tools, policies, budgets, and approval gates.

## 1\. Define the workflow job

Start with the outcome, not the models: support triage, document Q&A, escalation, classification, tool execution, or multi-step analysis.

## 2\. Connect model providers

Add the providers you need, such as Anthropic for reasoning, OpenAI or Azure OpenAI for general chat, Voyage AI for embeddings, Bedrock for AWS-governed access, or Groq/Together where supported.

## 3\. Create one deployment per model role

Create separate provider-backed deployments for each role, for example planner, classifier, summarizer, and final answer model. Each deployment can have its own provider, contract, policies, and budget gates.

## 4\. Create the workflow graph

Use the workflow editor to compose supported nodes such as input, retrieval, mcp\_input, tool\_call, function, model, agent, router, condition, approval, and output. Each model or agent node can bind to a different provider-backed deployment.

## 5\. Add routing and fallbacks

Use router or condition nodes for tiering and escalation. Use fallback\_node\_id for supported fallback behavior when a node fails.

## 6\. Add tools and MCP inputs

Use function/tool\_call nodes for HTTP/webhook/internal tools, and mcp\_input nodes for external MCP servers. Agent nodes can expose those functions as native provider tools.

## 7\. Put policy boundaries before risky steps

Add schema contracts, redact\_text, policy\_check nodes, budget policies, approval nodes, and enforcement packs before expensive or sensitive model calls.

## 8\. Bind deployments to model and agent nodes

In the workflow editor, select each model or agent node and choose the deployment that matches its role. You can also use a run default deployment for simple graphs.

## 9\. Run and inspect the trace

Execute the workflow, then inspect node executions, tool calls, model payloads, latency, cost, policy decisions, and errors in observability.

## 10\. Add evals for critical paths

Create golden datasets for the workflow’s important tasks. Use eval runs and comparisons to catch regressions after prompt, provider, routing, or policy changes.

## Recommended graph pattern

Start simple, then add routing and fallbacks once the main path is reliable:

```
input
  → redact_text / policy_check
  → planner model
  → retrieval or mcp_input
  → agent or function
  → router / condition
  → specialist model
  → approval (optional)
  → final answer model
  → output
```

Bind each `model` or `agent` node to the deployment that fits the role. For example, use a cheaper classifier model for routing and a stronger reasoning model for native tool-calling.

## Common questions

### Can every model node use a different provider?

Yes. Norcaster workflows support per-model-step deployment bindings, so one graph can use different providers and models for different roles.

### Does Norcaster execute agent nodes today?

Yes. Agent nodes run a bounded native tool-calling loop against a provider-backed chat deployment. Tools are configured as HTTP, webhook, or internal functions and every step is recorded in the workflow trace.

### How do I prevent runaway workflow cost?

Use budget policies, retry limits, route-specific deployments, and approval gates before expensive or tool-heavy branches.

### How do I audit what happened?

Each workflow execution records node-level inputs, outputs, errors, timings, traces, provider/model metadata, and policy decisions.