smart-delegation

Smart Delegation

Route tasks to the right agent at the right model tier. Stop wasting Opus tokens on file searches and Haiku tokens on architecture decisions.

Announce: "Using smart-delegation to route this task."

Quick Start

Before starting any task, score it:

Score >= 3 → Delegate to subagent
Score < 3  → Handle it yourself

Scoring (add up what applies):

Factor	Points
Touches 3+ files	+2
Bulk/repetitive operation	+2
Research or learning task	+2
Code review	+2
Exploration or search	+3
Independent (no shared state)	+2

Subtract:

Factor	Points
Contains: production, deploy, secret, password	-10
User said "show me" or "explain"	-5

Score >= 3? Delegate. Move on.

The Full System

Step 1: Score the Task

Run the scoring table above. If the score is negative or zero, do it yourself.

Step 2: Pick the Model

Match complexity to model. Don't overthink this.

Complexity	Model	Use For
1-2	Haiku	File search, pattern finding, simple lookups
3-6	Sonnet	Bug fixes, research, planning, code review
7+	Don't delegate	Deep architecture, multi-system reasoning

Why not delegate 7+? Complex tasks need your full context window. Subagents start fresh with zero context - they'll miss critical connections.

Step 3: Pick the Agent Type

Task	Agent Type	Why
Find files, explore code	Explore	Built for fast codebase search
Debug failing tests	debugger	Structured hypothesis testing
Plan implementation	Plan	Architecture-first thinking
Everything else	general-purpose	Full tool access

Step 4: Write the Prompt

Every delegation prompt needs 6 sections. No exceptions.

## 1. TASK
What to do. One sentence. Be specific.

## 2. EXPECTED OUTCOME
What the agent should return. Format, structure, length.

## 3. REQUIRED TOOLS
Which tools the agent should use. Be explicit.

## 4. MUST DO
Non-negotiable requirements. List each one.

## 5. MUST NOT DO
Guardrails. What the agent must avoid.

## 6. CONTEXT
File paths, patterns, constraints the agent needs to know.

Why 6 sections? Subagents have zero context about your project. The prompt IS their entire world. Vague prompts produce vague results.

Step 5: Set Effort Level

Effort	max_turns	When
Low	3	File search, quick lookups
Medium	6	Standard tasks (default)
High	12	Thorough review, edge cases
Max	25	Security audit, full architecture

Step 6: Verify the Result

After every delegation, check:

• Does the result match EXPECTED OUTCOME?
• Were all MUST DO items addressed?
• Were all MUST NOT DO items respected?

No verification = task not complete. Trusting without checking is the #1 delegation failure.

When to Use

• Task scores >= 3 on the scoring table
• You're running low on context window (delegation saves ~90% of tokens vs doing it yourself)
• Multiple independent tasks can run in parallel
• The task is well-defined enough to describe in 6 sections

When NOT to Use

• Task involves secrets, credentials, or production systems
• User explicitly wants to see the process ("show me", "walk me through")
• Task requires deep cross-file reasoning with shared state
• You can't clearly define the expected outcome (if you can't describe it, don't delegate it)

Examples

Example 1: Delegating a Code Search

Score: Exploration (+3) + Independent (+2) = 5. Delegate.

## 1. TASK
Find all API endpoint definitions in the codebase.

## 2. EXPECTED OUTCOME
List of endpoints with: HTTP method, path, handler file, line number.

## 3. REQUIRED TOOLS
Grep, Glob, Read

## 4. MUST DO
- Search for route definitions (app.get, app.post, router.*, @app.route, etc.)
- Include both REST and GraphQL endpoints
- Report file path and line number for each

## 5. MUST NOT DO
- Don't modify any files
- Don't execute any code

## 6. CONTEXT
Project uses Next.js with API routes in src/app/api/

Agent: Explore | Model: Haiku | Effort: Low (3 turns)

Example 2: Delegating a Bug Investigation

Score: Exploration (+3) + 3+ files (+2) + Independent (+2) = 7. Delegate.

## 1. TASK
Investigate why the email signup form returns 500 on submit.

## 2. EXPECTED OUTCOME
Root cause analysis with: error location, why it fails, suggested fix.

## 3. REQUIRED TOOLS
Read, Grep, Glob, Bash (for checking logs only)

## 4. MUST DO
- Read the form component and its submit handler
- Read the API route that handles the submission
- Check for recent changes to related files (git log)
- Identify the exact line where the error originates

## 5. MUST NOT DO
- Don't modify any files
- Don't run the application
- Don't access any external services

## 6. CONTEXT
Form component: src/components/email-capture.tsx
API route: src/app/api/subscribe/route.ts
Error started after last deploy (2 days ago)

Agent: debugger | Model: Sonnet | Effort: Medium (6 turns)

Example 3: Parallel Delegation

Three independent tasks? Launch them simultaneously:

Task A: Search for unused imports     → Explore, Haiku, Low
Task B: Review auth middleware        → general-purpose, Sonnet, High
Task C: Find test coverage gaps       → general-purpose, Sonnet, Medium

All three run in parallel. You get results from all three before deciding next steps.

Anti-Patterns

Pattern	Problem	Fix
"Just look into this"	No expected outcome	Define what "done" looks like
Delegating with full context	Agent gets 200K tokens of noise	Give only relevant context
Not verifying results	Agent may hallucinate or miss things	Always run Step 6
Delegating 7+ complexity	Agent lacks your mental model	Do it yourself
Sequential when parallel works	Wastes time	Launch independent tasks together

Security Guardrails

Never delegate tasks involving:

• API keys, tokens, passwords, secrets
• Production database operations
• Deployment or release processes
• User data access or PII handling
• Payment or financial operations

These require your direct attention and the user's explicit approval.

With Starter System

If you have the Starter System installed:

• Use /remember to log delegation patterns that work well
• Check /context-stats before delegating - high context usage makes delegation more valuable
• Use /handoff to pass delegation insights to the next session

With Kairn (Course)

When Kairn is configured, delegation gains:

• Adaptive scoring - learns which tasks delegate well and which don't from your history
• Outcome tracking - records delegation success/failure to refine routing over time
• Cross-session patterns - "Last 5 times you delegated auth reviews to Sonnet, 4 succeeded"

Without Kairn: scoring is manual and each session starts fresh. Still valuable, just not adaptive.