Decision: Design multi-agent architecture

[buddha314]

Decision Required

Design the FastAgent multi-agent system architecture for coordinating research, analysis, summary, and recommendation agents.

Key Decisions

1. Agent Roles and Responsibilities

• Research Agent (query understanding, multi-source search)
• Analysis Agent (trends, word clouds, keyword graphs)
• Summary Agent (article summarization, insights)
• Recommendation Agent (related papers, gap identification)
• Coordination Agent (orchestration)

2. Communication Patterns

• How do agents communicate with each other?
• Message passing vs shared state
• Event-driven vs request-response

3. State Management

• Centralized vs distributed state
• Conversation context management
• Workspace and document state

Impact

• Agent intelligence capabilities
• System complexity
• Scalability

Blocking

Server implementation (Phase 1)

References

• specs/TASKS.md (Phase 0, lines 18-22)
• specs/PROJECT_STATUS.md (lines 98-102)

[buddha314]

Decision Made: Event-Driven Coordinator Pattern

After comprehensive analysis, the recommended multi-agent architecture is an event-driven coordinator pattern with specialized agents.

Architecture Overview

Agent Coordinator (Core)

• Central request router and orchestrator
• Manages agent lifecycle and health
• Event bus (Redis pub/sub) for communication
• Task queue (Celery) for long-running operations

Specialized Agents

1. Research Agent - Query understanding, multi-source search (vector DB + MCP servers)
2. Analysis Agent - Trend analysis, word clouds, keyword graphs
3. Summary Agent - Article summarization using LLM
4. Recommendation Agent - Related papers, gap identification

Communication Pattern

• Event Bus: Redis pub/sub for real-time agent communication
• Task Queue: Celery for reliable long-running task execution
• State Management: Centralized Redis store for active tasks
• Persistence: PostgreSQL for long-term data

Multi-Agent Workflows

Agents can work independently or in coordinated workflows:

• Sequential: Research → Analysis → Summary → Recommendations
• Parallel: Multiple agents processing simultaneously
• Reactive: Agents respond to events from other agents

Key Benefits

1. Scalability - Agents scale independently via horizontal scaling
2. Resilience - Failure isolation, circuit breakers, retry logic
3. Flexibility - Easy to add new agent types
4. Real-time - Natural fit with WebSocket updates
5. Testability - Independent agent testing
6. Maintainability - Clear separation of concerns

Technology Stack

• Framework: FastAPI (REST + WebSocket)
• Event Bus: Redis pub/sub
• Task Queue: Celery
• Agents: Custom FastAgent implementation (lightweight)
• Monitoring: Prometheus metrics, structured logging

Documentation

Full architecture specification with code examples: specs/MULTI_AGENT_ARCHITECTURE.md

This decision unblocks Phase 1 server implementation.

Status: ✅ Decision finalized

[buddha314]

Decision documented and approved. Proceeding with event-driven coordinator pattern for multi-agent architecture.