Skip to main content

​
Project Goal

Build a production-grade multi-agent research system that demonstrates all major Module 5 concepts:
  • ✅ Workflow orchestration with parallel execution and quality loops
  • ✅ Agent orchestration with intelligent task decomposition
  • ✅ Hybrid architecture combining both patterns
  • ✅ Cost management with model cascading and iteration limits
  • ✅ Full observability with structured logging and tracing

​
System Architecture

User Query: "Research AI agent adoption in healthcare"
    |
    V
+-------------------------------------------------+
│ Coordinator Agent (Agent Orchestration)         │
│ - Analyzes query complexity                     │
│ - Decides: simple lookup OR multi-step research │
│ - Routes to appropriate workflow                │
+-------------------------------------------------+
              |
              V
    +-------------------+
    |                   |
    V                   V
[Simple Path]      [Complex Path]
Quick answer        Multi-agent workflow
                          |
                          V
              +--------------------------+
              │ Research Workflow        │
              │ (Workflow Orchestration) │
              +--------------------------+
                          |
                          V
              +-----------------------+
              |                        |
              V                        V
        [Parallel Research]      [Sequential Synthesis]
        - Market data            1. Combine findings
        - Academic papers        2. Quality check
        - Industry reports       3. Generate report
        - Expert interviews      (Loop if quality < 80%)
              |
              V
        Quality Gate → Output

​
Project Requirements

1. Workflow Orchestration Implement multi-agent workflow with:
  • A. Parallel Research Stage
  • B. Sequential Synthesis with Quality Loop
Must demonstrate:
  • 4 agents execute in parallel during research phase
  • Latency improvement: parallel vs. sequential (measure both)
  • Quality loop with iteration limit (max 3 iterations)
  • Early termination when quality threshold met (>= 80 score)
  • Cost tracking per agent and per phase
2. Agent Orchestration Implement coordinator agent that intelligently routes between simple queries and complex queries.
  • SIMPLE QUERIES (route to quick_answer)
    • Factual lookups (“What is X?”)
    • Simple definitions
    • Single data points
  • COMPLEX QUERIES (route to multi_agent_workflow)
    • Market analysis
    • Comparative research
    • Trend analysis
    • Multi-faceted topics
Must demonstrate:
  • Coordinator correctly classifies query complexity (test 10+ queries)
  • Simple queries route to single agent (< 2 sec latency)
  • Complex queries route to multi-agent workflow (6-10 sec latency)
  • Cost comparison: simple path (0.01)vs.complexpath(0.01) vs. complex path (0.15)
  • Routing accuracy: >= 85% correct classification
3. Hybrid Architecture Combine agent and workflow orchestration Must demonstrate:
  • Clear separation: agent for routing, workflow for execution
  • Justification for each choice (why agent here, workflow there)
  • Performance comparison vs. pure agent orchestration
  • Cost comparison vs. pure agent orchestration
  • Reliability comparison (determinism where possible)
4. Production Patterns Implement essential production features:
  • A. Cost Management
  • B. Observability
  • C. Delegation Loop Prevention
  • D. Timeout Protection
Must implement:
  • Cost tracking with budget limits
  • Structured logging with request IDs
  • Delegation tracking prevents loops
  • Timeout protection on all agents
  • Circuit breaker for repeated failures
  • Metrics dashboard (latency, cost, success rate)

​
Bonus Challenges

Choose one or more:
  • A2A Integration: Expose one specialist agent via A2A, consume from coordinator
  • Advanced Parallelism: Implement batching for 100 queries simultaneously
  • Dynamic Tool Selection: Coordinator chooses tools based on query type
  • Multi-Framework: Use both LangGraph and Google ADK in same system
  • Streaming Results: Stream partial results as agents complete
  • Human-in-the-Loop: Add approval gate for expensive operations (> $0.20)
  • Adaptive Budgets: Allocate more budget for complex queries automatically

​
Metrics to Track

Workflow Orchestration:
  • Parallel execution latency vs. sequential baseline
  • Quality loop: avg iterations, max iterations hit rate
  • Cost per workflow stage
  • Target: 60%+ latency reduction with parallel, < 2.5 avg iterations
Agent Orchestration:
  • Routing accuracy (simple vs. complex classification)
  • Cost savings from intelligent routing (simple path < 0.02,complex<0.02, complex < 0.20)
  • Latency by path (simple < 3 sec, complex < 12 sec)
  • Target: 85%+ routing accuracy, 80%+ cost savings on simple queries
Hybrid Architecture:
  • Overall system latency (p50, p95, p99)
  • Overall system cost per query (by complexity)
  • Success rate (queries completing without errors)
  • Target: p95 < 15 sec, avg cost < $0.10, success rate > 95%
Production Patterns:
  • Budget exceeded rate (should be < 1%)
  • Timeout rate (should be < 2%)
  • Delegation loops prevented (should be 0)
  • All requests have unique request IDs and full trace logs

​
Resources