AI Agent Swarm self-learning multi-agent collaboration system design

You are an expert AI systems architect specializing in multi-agent swarm intelligence. Design a self-learning multi-agent swarm collaboration system with the following specifications: ## Architecture Requirements 1. **Agent Registry & Discovery**: Design a service mesh where agents can register capabilities, discover peers, and negotiate task assignments 2. **Communication Protocol**: Define an inter-agent message format supporting: - Task delegation with priority levels - Partial result streaming between agents - Conflict resolution when multiple agents claim the same subtask 3. **Shared Memory Layer**: Design a memory architecture with: - Short-term working memory (per-task context) - Long-term knowledge base (learned patterns across sessions) - Episodic memory (successful/failed execution traces) 4. **Self-Evolution Mechanism**: - After each task completion, agents evaluate their performance - Successful strategies are promoted to shared skill library - Failed approaches are logged with root cause analysis - Agent specialization emerges from repeated task exposure ## Deliverables - System architecture diagram (describe in Mermaid syntax) - Agent communication protocol specification - Memory schema design - Self-evolution feedback loop pseudocode - Deployment topology for 5-50 concurrent agents ## Constraints - Must work with heterogeneous LLM backends (GPT-4, Claude, local models) - Latency budget: <2s for agent-to-agent communication - Must support graceful degradation when agents fail - Total token budget awareness and cost optimization routing Provide the complete system design with code examples where applicable.