Context Engineering

Overview

State-of-the-art patterns for managing context in LLM agent systems. These patterns enable complex multi-agent workflows while minimizing token overhead through strategic context engineering.

The Four Laws of Context Management

Law	Principle	Token Impact
1. Selective Projection	Pass only fields each agent needs	-30-50%
2. Tiered Fidelity	Define explicit context tiers per role	-40-60%
3. Reference vs Embedding	Use references for large data	-50-80%
4. Lazy Loading	Load data on-demand, not upfront	-30-50%

For detailed explanations and examples, see references/four-laws.md.

Context Tiers

Tier	Description	Use Case	Typical Size
FULL	Complete data	Initial analysis	5-20K tokens
SELECTIVE	Relevant subset	Domain workers	1-5K tokens
FILTERED	Criteria-matched	Validators	500-2K tokens
MINIMAL	Mode + counts	Routing	100-500 tokens
METADATA	Stats only	Synthesis	50-200 tokens

For tier selection guidance, see references/context-tiers.md.

Quick Reference: Input Section Pattern

Before (Anti-pattern)

## Input
You receive:
- snapshot: Full context snapshot
- all_findings: Complete list
- full_config: Everything

After (SOTA Pattern)

## Input
You receive (SELECTIVE context):
- analysis_summary: Key findings only
- relevant_files: Files for this focus area
- mode: Analysis depth setting

**NOT provided** (context isolation):
- Full plugin contents
- Unrelated analysis results
- Other agents' intermediate work

Anti-Patterns to Avoid

Anti-Pattern	Problem	Fix
Snapshot Broadcasting	Same data to every agent	Tier by role
Defensive Inclusion	"Maybe they need this"	Document NOT PASSED
Grounding Everything	Validating low-priority	Severity batching
Large Embeddings	Full arrays when counts suffice	Reference pattern
Repeated Context	Same data multiple times in chain	Pass once, reference later

Handoff Protocol

Standard handoff between agents:

handoff:
  from_agent: coordinator
  to_agent: analyzer
  context_level: SELECTIVE

  payload:
    mode: deep
    analysis_summary:
      claim_count: 15
      high_risk_count: 4
    relevant_files:
      - file: "[path]"
        content: "[content]"

  not_passed:
    - full_snapshot
    - unrelated_files
    - other_agents_data

  expected_output:
    format: yaml
    schema: AnalysisOutput

For complete handoff patterns, see references/handoff-protocols.md.

Severity-Based Batching

Reduce validation operations by priority:

batching:
  HIGH:     [all_validators]    # 4 agents
  MEDIUM:   [checker, estimator] # 2 agents
  LOW:      [checker]            # 1 agent
  INFO:     []                   # Skip

# Result: 60-70% fewer validation operations

Metrics to Track

Metric	Target	Calculation
Tier Compliance	100%	Agents with tier / Total agents
Redundancy Ratio	< 0.1	Duplicate data / Total data
Context per Agent	< 2K	Avg tokens per agent
NOT PASSED Coverage	100%	Agents with exclusions / Total

Additional Resources

• references/four-laws.md - Detailed law explanations with examples
• references/context-tiers.md - Tier definitions and selection guide
• references/handoff-protocols.md - YAML schema patterns
• references/examples.md - Production examples from red-agent