Smart routing layer that analyzes tasks and recommends the optimal agent to minimize Opus usage while maintaining quality.
When to Use
Always call analyze_task_routing before spawning a sub-agent. This prevents defaulting to Opus for tasks that Sonnet or Haiku can handle.
Tool: analyze_task_routing
Analyzes a task and returns the recommended agent with reasoning.
Usage:
// Before spawning
const routing = await analyze_task_routing({
task: "Polish the analytics dashboard UI",
context: "React + Tailwind, needs to match Command Center aesthetics"
});
// Then spawn with the recommendation sessions_spawn({ agentId: routing.agent, task: task_description, // ... other params });
## Routing Criteria
### Haiku (DEFAULT for Simple Tasks — Feb 2, 2026 Update)
**Use for:**
- Simple file edits (formatting, renaming, basic refactoring)
- Quick Q&A or fact lookups
- Basic data transformations
- Status checks, summaries under 1000 words
- Repetitive tasks with clear instructions
- Single-file research summaries
- Web scraping with clear targets
- File organization/moving
- Basic API calls with known patterns
- Generating lists or enumerations
- Simple templating tasks
**Try Haiku FIRST if the task:**
- Has clear, specific instructions
- Doesn't require creative judgment
- Can be validated objectively
- Follows a known pattern
**Escalate to Sonnet only if:**
- Task requires reasoning or judgment
- Multi-step workflows with dependencies
- Output quality needs to be nuanced
### Sonnet (Default Workhorse)
**Use for:**
- General coding tasks (features, bug fixes, testing)
- Research and analysis
- Content generation (blog posts, docs, reports)
- Data processing and transformation
- Most API integrations
- File reorganization with logic
- Standard debugging
**This should handle 70-80% of tasks.**
### Codex (Reasoning for Code)
**Use for:**
- Complex UI/UX implementation
- Architecture decisions and refactoring
- Debugging subtle/tricky bugs
- Performance optimization
- Code reviews requiring deep analysis
- Building entire features from scratch
- Complex algorithm implementation
**Codex uses o1 reasoning - great for "thinking through" code problems.**
### Opus (Strategic Oversight)
**Reserve for:**
- Multi-agent coordination
- Complex strategic decisions spanning multiple domains
- Sensitive business logic requiring judgment
- Tasks requiring McKinzie's context/preferences
- Novel problems without clear solutions
- High-stakes decisions
**Opus should be rare - most work should delegate down.**
## Implementation Notes
- **Routing uses Sonnet** to analyze tasks (cheap, fast, smart enough)
- **Default to Sonnet if uncertain** - safer than Haiku, cheaper than Opus
- **I can override** the recommendation if context requires it
- **Track routing decisions** in session notes for improvement
## Examples
```typescript
// Simple edit → Haiku
analyze_task_routing({ task: "Fix indentation in config.json" })
// → { agent: "haiku", reasoning: "Simple formatting task" }
// Feature work → Sonnet
analyze_task_routing({ task: "Add email validation to signup form" })
// → { agent: "sonnet", reasoning: "Standard feature implementation" }
// Complex UI → Codex
analyze_task_routing({ task: "Build interactive data visualization with D3" })
// → { agent: "codex", reasoning: "Complex UI requiring architectural decisions" }
// Strategic → Opus
analyze_task_routing({ task: "Design routing strategy for multi-agent system" })
// → { agent: "opus", reasoning: "Meta-level strategic decision" }
Cost Impact
Typical session costs (rough estimates):
- Haiku: $0.01 - $0.10
- Sonnet: $0.10 - $1.00
- Codex: $0.50 - $5.00 (reasoning tokens)
- Opus: $2.00 - $20.00
Using this routing can save 60-80% on agent costs by preventing Opus from handling Sonnet-tier work.