Magentic

Magentic is a better Facebook marketplace where bidding, negotiating, and offers are done by agents for both buyers and sellers. Set parameters, and let your agent do the work for you.

Comment 1

Inspiration

We were frustrated by the broken experience of online marketplaces like Facebook Marketplace. Buyers and sellers waste countless hours messaging back and forth, only to be ghosted at the last minute. Deals that should take days drag on for months, and good opportunities slip away because someone didn't respond fast enough. We imagined a world where AI agents could negotiate on our behalf - instantly, intelligently, and without the human flakiness that ruins online transactions.

What it does

Magentic is an AI-powered marketplace where autonomous agents negotiate deals on behalf of buyers and sellers. When a buyer finds a product they want, they simply set their maximum price and let their AI agent take over. The buyer agent communicates with the seller agent in real-time, asking pre-check questions about product condition, making strategic offers, and negotiating to find a mutually acceptable price, all without human intervention. The agents coordinate safe meetup locations and times, alerting humans only at critical checkpoints for final approval. The result? Deals close 10x faster with zero ghosting.

How we built it

Frontend: Built with Next.js 16 and TypeScript, styled with Tailwind CSS 4.0 for a modern magazine-aesthetic design. We used Framer Motion for smooth animations and Radix UI for accessible components. AI Agent Infrastructure: Deployed buyer and seller agents on Fetch.ai's Agentverse platform. The agents communicate via the uAgents protocol and use Claude AI (accessed through Lava Gateway) for intelligent decision-making during negotiations. Lava Gateway optimizes our Claude API calls, reducing latency and costs.

Agent Intelligence: Each agent has its own negotiation strategy:

  • Buyer Agent: Starts at ~91% of the user's max price, asks condition questions, negotiates strategically upward
  • Seller Agent: Starts at ~96% of listing price, responds to inquiries, makes counter-offers to maximize profit while staying above their floor price
  • Real-time Chat: The frontend displays live agent-to-agent conversations, showing the entire negotiation transcript as messages stream in. Users can watch their agents work and approve final deals.

Challenges we ran into

Agent Communication Architecture: Initially tried to call Agentverse agents via REST endpoints, but learned agents communicate via uAgents protocol, not simple HTTP calls. We adapted by implementing agent logic that mirrors our deployed agents' behavior for demo reliability. Real-time Message Streaming: Coordinating asynchronous agent responses and displaying them in real-time required careful state management and callback handling. Negotiation Logic: Designing agents that negotiate intelligently (not too aggressive, not too passive) required multiple iterations and strategy testing to find what works. Lava Gateway Integration: Learning to route Claude API calls through Lava Gateway for optimized performance while maintaining fast response times.

Accomplishments that we're proud of

  • Fully Functional Agent Negotiation: Our agents successfully negotiate deals, reaching mutually acceptable prices in 4-6 message rounds
  • Intelligent Pre-checks: Agents ask relevant questions about product condition before making offers
  • Strategic Pricing: Agents use market context and user constraints to make smart decisions
  • Seamless UX: Built a beautiful, magazine-style interface that makes complex agent interactions feel simple
  • Autonomous Coordination: Agents handle everything from price negotiation to meetup scheduling without human micromanagement

What we learned

Agent architecture on Fetch.ai: Understanding the uAgents protocol and how to deploy persistent agents on Agentverse

LLM routing optimization: How Lava Gateway can dramatically improve AI API performance and cost-efficiency

Autonomous negotiation strategies: Designing AI behavior that feels natural and achieves win-win outcomes

Real-time state management: Handling asynchronous agent messages and keeping the UI synchronized

User trust in AI: The importance of showing users what agents are doing rather than hiding the process in a black box

What's next for Magentic

Enhanced Agent Capabilities: Add multi-item bundle negotiations, price history analysis, and reputation-based trust scoring

Escrow Integration: Implement secure payment and escrow services for complete transaction safety

Mobile App: Build React Native apps for iOS and Android with push notifications for deal updates

Multi-Marketplace Support: Expand beyond our platform to negotiate/intake data from Facebook Marketplace, Craigslist, and OfferUp

Advanced AI Features: Implement computer vision for product verification and fraud detection

TEMP VID CAUSE WIFI

Built With

Share this project:

Updates