MCP Buyer seller broker

What it does The MCP Buyer-Seller-Broker is a decentralized AI agent system that enables agents to autonomously negotiate, transact, and make decisions on behalf of humans. Agents represent buyers, sellers, and brokers—each with their own goals, preferences, and memory. Using Model Context Protocol (MCP) and Claude 3.5 Sonnet via AWS Bedrock, the agents share context, evaluate options, and collaboratively reach agreements. Our initial use case is ticket brokering, but the architecture supports any transactional flow.

🛠️ How we built it We used a fully serverless AWS stack:

Amazon API Gateway: Entry point for agent communication

AWS Lambda (Python 3.13): Agent logic, prompting, memory fetch

DynamoDB: AgentRegistry and ProductRegistry

Amazon S3: Storage for FAISS vector index (semantic memory)

FAISS (via Lambda Layer): Vector search for memory/context retrieval

Titan Embeddings v2: Embeds agent profiles and offers

Claude 3.5 Sonnet via Bedrock: Core LLM for decision-making

Jupyter Notebook: For local agent simulation and development

Agents follow structured prompting via MCP, which defines shared context schemas and interaction flow between agent types.

🧗 Challenges we ran into Cold start limits for Lambda with FAISS

Prompt stability with Claude across agent types

Ensuring negotiation flows don’t loop or stall

Memory vectorization and retrieval accuracy

Designing MCP as both a communication protocol and logic layer

Orchestrating multi-agent coordination in real time

🏆 Accomplishments that we're proud of Created a functional MCP agent bridge with real memory and vector search

Enabled Claude to simulate multi-agent negotiation with persistent context

Built a full buyer-seller-broker chain using only AWS serverless resources

Maintained modularity so new agent roles can be added with minimal effort

Reduced cost and complexity by leveraging Bedrock and Titan embeddings

📚 What we learned Structured prompts with embedded memory yield far better results

Serverless design scales beautifully—but requires thoughtful cold start management

Claude 3.5 Sonnet is strong at reasoning, but semantic clarity is critical

Vector search is invaluable for maintaining long-term memory across agents

Defining interaction protocols (MCP) creates reusable, interpretable systems

🔮 What's next for MCP Buyer-Seller-Broker Support multi-party negotiations (e.g., auctions, bundles, add-ons)

Add agent personality traits and risk profiles

Move toward real-time chat interface with human input

Integrate event-driven workflows via AWS EventBridge or Step Functions

Expand to new verticals like real estate, freelance hiring, and logistics

Publish a developer SDK for third-party AI agent integration

Built With

• amazon-web-services
• claude
• faiss
• lambda
• titan