Skip to content

🤖 [VISION - Not MVP] Multi-Agent AI Architecture

Timeline: Post-MVP, triggered by performance metrics Current Status: Concept only Warning: Do not implement during MVP phase

Overview

Evolution from single agent to specialized multi-agent system following Anthropic's guidance on building effective agents.

Evolution Triggers

Implement multi-agent when:

  • Average latency exceeds 2 seconds
  • Request queue depth > 100
  • Clear workflow patterns emerge from logs
  • Team has dedicated AI engineer

Architecture Vision

Phase 1: Current MVP (Single Agent)

class GetCimpleAgent:
    """One agent handling all workflows"""
    def handle(self, request):
        # Route to appropriate workflow
        # Single point of failure/optimization

Phase 2: Four Specialized Agents

User Request
    ↓
Orchestrator Agent (Router)
    ↓
┌─────────────┬──────────────┬───────────────┐
Compliance    Action         Knowledge       Support
Agent         Agent          Agent           Agent

Agent Specializations

Orchestrator Agent

  • Request routing and prioritization
  • Response coordination
  • Session management
  • Error recovery

Compliance Agent

  • Essential Eight assessments
  • Risk evaluations
  • Framework mapping
  • Evidence validation

Action Agent

  • Task generation
  • Notification dispatch
  • Workflow execution
  • Integration triggers

Knowledge Agent

  • Question answering
  • Report generation
  • Pattern recognition
  • Insight synthesis

Implementation Principles

Following Anthropic's recommendations:

  1. Workflow > Prompts: Build deterministic workflows, use LLM only where needed
  2. Augment Don't Replace: Enhance human capability, don't try to replace it
  3. Simple Primitives: Combine simple, reliable components
  4. Extensive Logging: Track everything for continuous improvement

Performance Targets

Metric Single Agent (MVP) Multi-Agent (Vision)
Latency < 2s < 500ms
Throughput 100 req/min 1000 req/min
Accuracy 90% 95%+
Uptime 99% 99.9%

Resource Requirements

  • Team: 2 dedicated AI engineers
  • Infrastructure: Distributed compute
  • Timeline: 6-month implementation
  • Investment: [Post-Series A]

Risk Mitigation

  • Start with async agent communication
  • Build comprehensive fallback system
  • Maintain single-agent mode for DR
  • Extensive A/B testing rollout

Success Metrics

  • Response time improvement: >75%
  • Support ticket reduction: >50%
  • User satisfaction increase: >30%
  • System reliability: >99.9%

References

Decision Record

See /docs-internal/docs/02-strategy/decision-log.md entry:

  • 2025-01-11: Single Agent Architecture for MVP

Remember: This is a vision document. Do not implement until evolution triggers are met.