Multi-Model Mode

Use multiple LLMs in a single request. Reasoning model thinks, tool model executes, synthesis model responds.

---

Why multiple models?

Different models are good at different things. Opus is great at reasoning but expensive. Haiku is fast but shallow. GPT-4o handles tools well. o1 does extended thinking.

Multi-model mode picks the right model for each phase of a request instead of making one model do everything.

---

The four roles

Role	What it does	Default model
Reasoning	Analyzes the problem, plans approach, extended thinking	Claude Opus, o1
Tool Execution	Calls MCP tools, executes functions	Claude Sonnet, GPT-4o
Synthesis	Combines results into final response	Claude Sonnet, GPT-4o
Fallback	Error recovery when something fails	Gemini Flash, Haiku

---

How it decides

Not every request needs multi-model. The system analyzes:

1. Query complexity - Simple questions go to a single model

2. Tool requirements - If tools are likely needed, tool model gets involved

3. Slider position - Higher slider = more likely to use multi-model

4. Trigger patterns - Certain keywords ("analyze", "investigate", "debug") trigger it

---

Handoffs

Models pass context to each other through handoffs. The reasoning model's analysis becomes context for the tool model. Tool results become context for synthesis.

Max handoffs are configurable (default: 5) to prevent infinite loops.

---

Configuration

Environment variables

# Enable multi-model
ENABLE_MULTI_MODEL=true

# Per-tier model assignments
MULTI_MODEL_ECONOMY_REASONING=claude-3-haiku
MULTI_MODEL_ECONOMY_TOOL=claude-3-haiku
MULTI_MODEL_BALANCED_REASONING=claude-3-5-sonnet
MULTI_MODEL_BALANCED_TOOL=gpt-4o
MULTI_MODEL_PREMIUM_REASONING=claude-opus-4
MULTI_MODEL_PREMIUM_TOOL=claude-3-5-sonnet

# Extended thinking budget (tokens)
MULTI_MODEL_PREMIUM_THINKING_BUDGET=16000

Runtime toggle

Admins can enable/disable via the Admin Portal under Pipeline Settings.

Via SDK

const agent = aw.createAgent({
  model: 'auto',
  multiModel: {
    enabled: true,
    reasoningModel: 'claude-opus-4',
    toolModel: 'gpt-4o',
    synthesisModel: 'claude-3-5-sonnet'
  }
});

---

Cost tracking

Each role tracks its own costs. The total is reported in the response metadata:

{
  "multiModel": {
    "rolesExecuted": ["REASONING", "TOOL_EXECUTION", "SYNTHESIS"],
    "handoffCount": 2,
    "costBreakdown": {
      "REASONING": 0.045,
      "TOOL_EXECUTION": 0.012,
      "SYNTHESIS": 0.008
    },
    "totalCost": 0.065,
    "totalDuration": 8500
  }
}

---

When to use it

Good for:

• Complex analysis tasks

• Multi-step tool workflows

• Architecture and debugging questions

• Anything where you'd want to "think then do"

Skip it for:

• Simple Q&A

• Single-tool operations

• High-volume, low-complexity tasks

---

Limitations

• Adds latency (multiple model calls)

• Higher cost for complex requests

• Handoff context has size limits

• Not all providers support extended thinking