all factors
IIfoundation

Context Loading

Load when needed, not all at once. Stay under 40%.

3 min read

II. Context Loading

II
The 40% Rule

Load context when needed, not all at once. Never exceed 40%.


The Problem

::: danger Context Collapse When context exceeds ~40%, quality degrades suddenly—not gradually. You'll see:

  • Inconsistent outputs
  • Forgotten instructions
  • Hallucinated details
  • Lost focus on goals :::

Monolithic Context

Main Agent: 80% full

  • Task goals — 5%
  • Codebase exploration — 25%
  • Research findings — 20%
  • Implementation details — 15%
  • Validation results — 10%

Result: Context collapse

JIT Loading

Main Agent: 15% full

  • High-level goals — 5%
  • Orchestration logic — 5%
  • Compressed summaries — 5%

Result: Clean, focused


The Solution

Sub-Agent Architecture

Main Agent

10-20% context
Orchestrates only

Explorer

Fresh context
Returns summary

Implementer

Fresh context
Returns diff

Validator

Fresh context
Returns pass/fail

::: tip Key Insight Each sub-agent gets a fresh context window. Deep work happens in isolation. Only compressed results flow back to main. :::


The 40% Rule

Context LevelStatusAction
0-30%OptimalContinue working
30-40%CautionConsider delegation
40-60%DegradedSpawn sub-agent
60%+CollapseStop, compress, restart

::: warning Why 40%? This threshold appears across cognitive systems: ADHD hyperfocus, LLM coherence, database performance. It's not arbitrary—it's empirical. :::


Implementation

::: code-group

# Instead of exploring in main context:
# Use Task tool to spawn explorer
Task(
  prompt="Find all auth-related files",
  subagent_type="Explore"
)
# Returns: 3-5 line summary, not full content
# Sub-agent returns compressed insight:
"Auth handled in src/auth/:
 - middleware.ts (JWT validation)
 - routes.ts (login/logout)
 - types.ts (User interface)"

# NOT the full 500-line files
# Research phase → fresh context
# Plan phase → fresh context
# Implement phase → fresh context
# Each starts clean

:::


Real Example

Session Start: 0% context

1. User: "Add rate limiting to API"

2. Main agent (5% context):
   → Spawns Explorer sub-agent

3. Explorer (fresh context):
   → Reads 15 files
   → Returns: "API routes in src/routes/api.ts,
     uses Express middleware pattern"
   → Context discarded

4. Main agent (10% context):
   → Has summary, not 15 files
   → Spawns Implementer

5. Implementer (fresh context):
   → Reads api.ts + rate-limit docs
   → Returns: diff + "Added express-rate-limit"
   → Context discarded

6. Main agent (15% context):
   → Clean, focused, successful

Checklist

  • Main agent stays under 20% context
  • Delegate exploration to sub-agents
  • Sub-agents return summaries, not full content
  • Fresh context for each workflow phase
  • Compress before crossing 40% threshold

FactorRelationship
I. Automated TrackingGit stores what context discards
III. Focused AgentsSub-agents do one job
VI. Resume WorkBundles restore context across sessions