Multi-Agent Collaboration

Overview

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

Research Foundation

• Google ADK: Context compilation pipelines and session management
• Anthropic: Multi-agent coordination and handoff protocols
• Progressive Disclosure: Agent-readable semantic interfaces
• LangGraph/CrewAI/AutoGen: Framework-specific orchestration patterns

Pattern Selection Framework

Pattern	Use When	Trade-offs
Hierarchical	Clear decomposition, audit trails	Central bottleneck, sequential latency
Swarm	Parallel exploration, diverse perspectives	Coordination overhead, emergent behavior
ReAct	Dynamic adaptation, tool-heavy workflows	Myopic decisions, may meander
Plan-Execute	Clear sequence, predictability needed	Less adaptive, requires replanning
Reflection	Quality refinement, self-correction	Added latency, may reinforce errors
Hybrid	Multiple coordination needs	Implementation complexity

For detailed YAML definitions and examples of each pattern, see references/patterns.md.

The Four Laws of Context Management

Law 1: Selective Projection

Pass only fields each agent needs, not full data structures.

# BAD: Full snapshot everywhere
snapshot: {...20KB...}

# GOOD: Selective projection
context:
  mode: deep
  claims_analyzed: 15
  high_risk_count: 4

Law 2: Tiered Context Fidelity

Define explicit tiers based on agent role:

Tier	Description	Example Agent
FULL	Complete data	Initial analyzer
SELECTIVE	Relevant subset	Domain workers
FILTERED	Criteria-matched	Validators
MINIMAL	Mode + counts	Strategy/routing
METADATA	Scope stats only	Report synthesis

Law 3: Reference vs Embedding

For large data, pass reference instead of full structure:

# Embedding (expensive)
raw_findings: [{...}, {...}, ...]  # 40+ items

# Reference (efficient)
findings_summary:
  total: 45
  by_severity: {CRITICAL: 3, HIGH: 12}
  # Agent fetches specific findings on-demand

Law 4: Lazy Loading

Load data on-demand, not upfront:

initial_context:
  scope: {item_count: 45}
  available_data:
    - name: findings
      fetch: "request by severity or ID"

For implementation details and patterns, see references/context-engineering.md.

Standard Handoff Protocol

handoff:
  from_agent: context-analyzer
  to_agent: attack-strategist
  context_level: MINIMAL

  payload:
    mode: deep
    analysis_summary:
      claim_count: 15
      high_risk_count: 4
      patterns: [pattern_1, pattern_2]

  expected_output:
    format: yaml
    schema: strategy_v1

Severity-Based Batching

Reduce validation operations by routing based on priority:

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

# Result: 60-70% fewer operations

Anti-Patterns to Avoid

1. Snapshot Broadcasting - Passing full context to every agent
2. Defensive Over-inclusion - "Maybe they need this" mentality
3. Grounding Everything - Validating low-priority items
4. Embedding Large Lists - Full arrays when counts suffice
5. Repeated Context - Same data passed multiple times in chain
6. Verbose Outputs - Over-explaining when concise suffices

Progressive Disclosure for Agents

Three-Level Loading

level_1_always_loaded:
  - skill_name
  - skill_description
  tokens: ~100

level_2_on_trigger:
  - main_skill_body
  - core_patterns
  - quick_reference_tables
  tokens: ~2000

level_3_on_demand:
  - detailed_references
  - extended_examples
  - implementation_guides
  tokens: as_needed

Guardrails and Validation

Output Validation Pattern

validation:
  hook: post_tool_use
  on_invalid:
    action: block_and_retry
    max_retries: 2
  on_valid:
    action: continue

Context Tier Enforcement

Document what each agent does NOT receive:

agent_context:
  receives:
    - analysis_summary
    - assigned_vectors

  not_provided:  # CRITICAL: Explicit exclusions
    - full_snapshot
    - other_agents_data
    - conversational_arc

Metrics

Track these to validate optimization:

Metric	Target
Total context passed	< 100KB
Redundancy ratio	< 0.1
Validation efficiency	> 3:1 findings/operations
Tier compliance	100%

Additional Resources

• references/context-engineering.md - Detailed context management patterns
• references/patterns.md - Architectural patterns with YAML definitions
• references/examples.md - Red-agent implementation examples