Inspiration

As AI agents evolve from simple chatbots into autonomous actors that can book services and purchase goods on our behalf, a critical trust gap emerges. How can we allow agents to discover merchants, negotiate prices, and execute transactions without exposing users to unauthorized spending, out-of-scope actions, or token replay attacks? AgentMesh Web Ultra was inspired by the need to build a secure, open, and verifiable foundation for Agent-to-Agent (A2A) commerce.

What it does

AgentMesh Web Ultra is a full-stack autonomous Agent-to-Agent commerce node that enables safe, policy-controlled transactions:

  • Agent Discovery (/.well-known/agent.json): Automatically fetches and parses machine-readable service manifests before executing any commerce action.
  • Autonomous Goal Routing & Negotiation: Evaluates natural-language user goals (such as "Fix my flat tire as cheap as possible") to select the optimal provider and negotiates a real-time 15% discount with the merchant server.
  • Policy & Scope Guardrails: Blocks out-of-scope or unauthorized requests before execution.
  • Human-in-the-Loop (HITL) & Auto-Approve Limits: Supports configurable Auto-Approve spending limits ($0 to $50) for frictionless micro-transactions while enforcing a 300-second time-bound manual approval token for higher amounts.
  • Cryptographic Security & Replay Protection: Defends against forged confirmation tokens (HTTP 403 Spoof Defense), tracks single-use nonces to block duplicate execution (HTTP 409 Replay Guard), and issues verifiable hmac_proof receipts.

How we built it

We built AgentMesh Web Ultra using Next.js 14 (App Router), TypeScript, and React for a fast, responsive full-stack architecture. We implemented RESTful Agent Execution Endpoints (/api/agent/execute) alongside standard /.well-known/agent.json discovery manifests, cryptographic HMAC signing for receipts, and nonce-tracking middleware for replay protection. The application is deployed live on Vercel.

Challenges we ran into

Designing a seamless balance between autonomous execution and strict user safety was our biggest challenge. Ensuring that confirmation tokens remain strictly single-use (HTTP 409 on replay) while keeping the Human-in-the-Loop approval flow intuitive and responsive required careful state and nonce validation on the backend.

Accomplishments that we're proud of

  • Successfully implementing end-to-end A2A discovery, negotiation, and cryptographic receipt generation (hmac_proof) in a clean, interactive interface.
  • Building built-in security simulation tools (Simulate Spoof 403, Test Replay Guard 409, and Test Car Guardrail) so judges and developers can test edge-case defenses with a single click.

What we learned

We gained deep hands-on experience in designing machine-readable agent manifests, implementing cryptographic token validation, and building deterministic guardrails for autonomous AI workflows.

What's next for AgentMesh Web Ultra

  • Integrating decentralized identity (DID) and multi-merchant federated discovery networks.
  • Adding real-world payment rail integrations with programmable spending policies and multi-agent escrow support.

Built With

Share this project:

Updates

Submission history