Subagents offload context-expensive work. The caller receives a digest, not the full trace. Default subagents are read-only (research, analyze, summarize). READ-WRITE subagents may modify state within declared scope.

When to Delegate

See the contract's Subagent Delegation Protocol for authoritative triggers.

Summary:

Clarification on 250KB:

• Applies to Read operations (full content enters context)
• Does NOT apply to Grep (only matches enter context)
• Typical sequence: search (cheap) → identify matches → measure size of files to read → apply threshold

Keep Inline When

• Task requires user interaction mid-execution
• Result interpretation needs full conversation history
• Content to read ≤250KB and processing depth ≤2 steps

Delegation Protocol

1. Assess Scope (if uncertain)

# Check total size of files to READ (not search)
stat --printf="%s\n" src/api/*.py | awk '{sum+=$1} END {print sum}'

2. Generate Brief

MODE: SUBAGENT
MODE: SUBAGENT READ-WRITE    ← only when objective requires state changes
GOAL: {{objective}}
CONTEXT: {{what caller already knows — no pre-analysis required}}
SCOPE: {{files, directories, or boundaries}}

Brief principles:

• Write from existing knowledge only — if investigation is needed to specify the goal, that IS the delegated work
• Vague goals are valid when they reflect genuine uncertainty
• Subagent has the contract — don't repeat rules in the brief

3. Dispatch

Use Task tool. Subagent inherits contract but operates in Subagent Mode (no external gates, compressed output). Default is read-only; READ-WRITE permits state modification with mandatory Intent Gate per action.

3b. Model Selection

The CLAUDECODESUBAGENT_MODEL env var sets the default model. Override it per-agent when the task doesn't need the default's reasoning:

Rule of thumb: if the brief could be expressed as a shell pipeline but delegation is used for context isolation, use haiku.

For the Subagent

When receiving a MODE: SUBAGENT brief:

1. Review ~/§BRANDGLOBALDIRNAME§/AGENT_TOOLS.md — MCP servers often provide efficient alternatives to manual tool chains
2. Work within scope, abort if insufficient
3. Return structured output (see below)

4. Review Output

RESULT: [success | partial | blocked | failed]
SUMMARY: [what was found/analyzed]
CONCERNS: [issues for caller review]
BLOCKERS: [if not success — what prevented completion]
DETAILED RESPONSE: [findings]

Review: Verify result addresses goal intent. Address concerns before proceeding.

5. Integrate

• Success: Use digest, continue main task
• Partial: Complete inline or re-delegate with narrower scope
• Blocked/Failed: Surface to user or try alternative approach

Common Delegation Patterns

Anti-Patterns

Don't delegate:

• State-modifying operations without READ-WRITE marker — default subagents are read-only
• Decisions requiring user input
• The debugging hypothesis loop (requires iterative testing with full context)

Exception — debugging research IS delegatable:

• "Find all call sites of function X"
• "Find working analogues for this pattern"
• "Check if error handling exists for case Y"

These support Bug Qualification / Pattern Analysis phases without owning the hypothesis loop.

Parallel Delegation

Independent research tasks can run simultaneously:

1. Identify independent subtasks
2. Generate briefs for each
3. Dispatch in parallel (multiple Task tool calls)
4. Await all results
5. Integrate digests

Dependency awareness: If task B depends on task A's output, dispatch sequentially.