I’ve been thinking about the "last mile" problem for AI agents. We have agents that can code, plan, and browse, but they are still economically "trapped." They can't independently pay for their own API calls, compute, or data without a human-in-the-loop providing a credit card.
To solve this, I’m building the A2A (Agent-to-Agent) System, an open-source infrastructure designed to turn agents into independent economic actors.
What’s under the hood?
Identity (a2trust): DID-based verifiable identity using @veramo/core. It allows agents to establish persistent reputations (EigenTrust) so they can trust each other without centralized gatekeepers.
Payments (a2pay): Built on ERC-7579 Smart Accounts. Agents can use Session Keys to execute transactions autonomously within specific constraints (time-limited, amount-capped, gas-abstracted).
Protocol (a2api): A marketplace layer that utilizes MCP (Model Context Protocol). Agents can discover services via machine-readable docs (llms.txt) and negotiate fees via standard interfaces.
Why this matters:
Most current agent payment solutions are just wrappers around human wallets. A2A aims to build a native "Agent Economy" where an agent can earn revenue from its tools and spend it to hire other agents, creating a truly autonomous swarm.
The Tech Stack:
TypeScript/Node.js,
Viem / Permissionless.js for smart account abstraction,
MCP SDK for inter-agent communication,
Base L2 for low-cost transactions,
I’d love to get your feedback on the architecture, especially on the security implications of delegating session keys to LLM-driven agents.
GitHub: https://github.com/swimmingkiim/a2a-project
Edit: to be clear, A2A does not include payment functions but it has some projects close to it that look like it’s trying to solve the same thing. A2P comes to mind from Google: https://cloud.google.com/blog/products/ai-machine-learning/a...