iterative-retrieval

# Iterative Subagent Retrieval Protocol

A systematic approach to gathering context through subagents that ensures sufficient information before proceeding. Instead of accepting single-pass results, the orchestrator evaluates sufficiency and asks follow-up questions until confident.

## Core Insight

**Orchestrators have semantic context that subagents lack.**

When you dispatch a subagent, you know the purpose behind the request—the decision or action the information will support. The subagent only knows the literal query. Single-pass summaries often miss key details because the subagent lacks this implicit context.

**Solution:** Evaluate every subagent return and ask follow-up questions before accepting it. Loop until sufficient.

## When to Use

- Spawning Explore agents for codebase research
- Gathering context before planning decisions
- Investigating unfamiliar code areas
- Researching external documentation
- Any context-gathering task where completeness matters

## When NOT to Use

- Simple, targeted queries with clear answers
- Time-critical operations where latency matters
- Queries where the first result is definitionally sufficient
- Follow-up operations on already-gathered context

## The 4-Phase Protocol

### Phase 1: Initial Dispatch

Before dispatching, explicitly define your context needs:

1. **Define PRIMARY OBJECTIVE** - What decision or action does this information support?
2. **Define INITIAL QUERIES** - Specific items to retrieve
3. **Dispatch with BOTH** - Include queries AND objective context in the prompt
4. **Store agent ID** - Returned by Task tool for potential resumption

```javascript
// Example: Explore agent for understanding auth system
const result = Task({
  subagent_type: "Explore",
  description: "Research auth system",
  prompt: `OBJECTIVE: Planning implementation of OAuth2 refresh tokens

  QUERIES:
  1. Where is authentication currently handled?
  2. What token formats are used?
  3. How are sessions managed?

  WHY: Need to