A Novel Hierarchical Multi-Agent System for Payments Using LLMs
A new 4-level agent architecture tackles the critical gap where AI assistants like Operator can't handle money.
A team of researchers has introduced HMASP (Hierarchical Multi-Agent System for Payments), a groundbreaking framework designed to enable AI agents to autonomously execute payment workflows—a capability notably absent from current popular assistants like OpenAI's Operator and Claude's Computer Use. The system, detailed in a paper accepted at PAKDD 2026, directly addresses a critical gap in the agentic AI landscape, where automation stumbles at tasks involving financial transactions. By providing a structured, end-to-end method, HMASP lays the technical foundation for truly autonomous AI that can manage real-world business operations involving money.
The HMASP architecture employs a four-level hierarchy of specialized agents, starting with a Conversational Payment Agent (CPA) that handles user requests and coordinates downstream tasks. It leverages architectural patterns like shared state variables and structured handoff protocols to ensure modular, decoupled execution across agents. Crucially, the system is model-agnostic, capable of running on both open-weight (like Llama 3) and proprietary LLMs (like GPT-4). This research demonstrates the technical feasibility of agentic payments and opens the door for future development where AI can manage invoicing, reimbursements, and vendor payments without human intervention, significantly expanding the scope of automated workflows.
- Proposes HMASP, a 4-level agent hierarchy (CPA, Supervisor, Routing, Process Summary) for payment workflows.
- First LLM-based multi-agent system to implement end-to-end agentic payments, a key limitation for current AI assistants.
- Uses modular patterns like shared state and handoff protocols, and works with both open and proprietary LLMs.
Why It Matters
It enables AI to finally automate financial tasks, moving assistants from conversation to actionable business operations.