Mixture of Agents (MoA)

MoA intelligently routes tasks to the right model tier based on complexity — saving cost on simple tasks while reserving powerful models for hard ones.

What is MoA

Instead of sending every request to the most expensive model, MoA analyzes the incoming task and assigns it to one of three model tiers:

Tier	Typical Use	Cost	Latency
fast	Formatting, simple lookups, one-liners	$	~200ms
medium	Multi-step reasoning, moderate code generation	$$	~1s
powerful	Complex architecture, long-horizon planning, debugging	$$$	~3s

MoA is opt-in — disabled by default. Enable it via config or environment variable.

Task Complexity Detection

MoA classifies tasks using heuristic rules applied to the user message and context:

Signal	Simple	Medium	Complex
Message length	< 50 chars	50–300 chars	> 300 chars
Tool calls needed	0–1	2–4	5+
Code generation	None	Snippet / function	Multi-file / architecture
Context turns	1	2–5	6+
Keywords	"what is", "list", "convert"	"refactor", "explain", "write a function"	"design", "debug", "migrate", "architect"

The final complexity score is a weighted sum. Thresholds are configurable.

Model Tier Configuration

Define which model backs each tier in .quark-agent.json:

json { "moa": { "enabled": true, "defaultTier": "medium", "tiers": { "fast": { "provider": "openai", "model": "gpt-4o-mini" }, "medium": { "provider": "openai", "model": "gpt-4o" }, "powerful": { "provider": "anthropic", "model": "claude-sonnet-4-20250514" } } } }

Or via environment variables:

env MOA_ENABLED=true MOA_DEFAULT_TIER=medium MOA_FAST_MODEL=gpt-4o-mini MOA_MEDIUM_MODEL=gpt-4o MOA_POWERFUL_MODEL=claude-sonnet-4-20250514

Custom Routing Rules

Override the heuristic classifier with explicit rules for specific patterns:

json { "moa": { "enabled": true, "rules": [ { "match": "sql", "tier": "fast" }, { "match": "test", "tier": "medium" }, { "match": "security", "tier": "powerful" } ] } }

Each rule has:

• match — a substring or regex matched against the user message
• tier — the model tier to use
• priority (optional) — higher priority rules win when multiple match (default: 0)

Rules are evaluated before the heuristic classifier. First match wins.

Example Usage

Programmatic

```ts import { createAgent } from "quark-agent";

const { agent } = await createAgent({ profile: "coding", moa: { enabled: true, defaultTier: "medium", tiers: { fast: { provider: "openai", model: "gpt-4o-mini" }, medium: { provider: "openai", model: "gpt-4o" }, powerful: { provider: "anthropic", model: "claude-sonnet-4-20250514" }, }, rules: [ { match: "sql", tier: "fast" }, { match: "debug", tier: "powerful" }, ], }, });

// Simple query → routed to "fast" tier await agent.run("What is the current timestamp?");

// Complex task → routed to "powerful" tier await agent.run("Debug why the payment service returns 502 intermittently"); ```

CLI

```bash

Enable MoA with default tiers

quark-agent --moa

Override the powerful tier

quark-agent --moa --powerful-model claude-sonnet-4-20250514 ```

Inspecting routing decisions

Set LOG_LEVEL=debug to see which tier was selected and why:

[moa] task classified as "complex" (score: 0.82) → routing to "powerful" (claude-sonnet-4-20250514) [moa] rule match: "debug" → forcing tier "powerful"