A Novel Hierarchical Multi-Agent System for Payments Using LLMs
| Authors | Joon Kiat Chua et al. |
| Year | 2026 |
| Field | AI / ML |
| arXiv | 2602.24068 |
| Download | |
| Categories | cs.MA, cs.CL |
Abstract
Large language model (LLM) agents, such as OpenAI's Operator and Claude's Computer Use, can automate workflows but unable to handle payment tasks. Existing agentic solutions have gained significant attention; however, even the latest approaches face challenges in implementing end-to-end agentic payment workflows. To address this gap, this research proposes the Hierarchical Multi-Agent System for Payments (HMASP), which provides an end-to-end agentic method for completing payment workflows. The proposed HMASP leverages either open-weight or proprietary LLMs and employs a modular architecture consisting of the Conversational Payment Agent (CPA - first agent level), Supervisor agents (second agent level), Routing agents (third agent level), and the Process summary agent (fourth agent level). The CPA serves as the central entry point, handling all external requests and coordinating subsequent tasks across hierarchical levels. HMASP incorporates architectural patterns that enable modular task execution across agents and levels for payment operations, including shared state variables, decoupled message states, and structured handoff protocols that facilitate coordination across agents and workflows. Experimental results demonstrate the feasibility of the proposed HMASP. To our knowledge, HMASP is the first LLM-based multi-agent system to implement end-to-end agentic payment workflows. This work lays a foundation for extending agentic capabilities into the payment domain.
Engineering Breakdown
Plain English
This paper addresses a critical gap in LLM agent automation: while models like OpenAI's Operator and Claude can handle many workflows, they fail at payment tasks end-to-end. The authors propose HMASP (Hierarchical Multi-Agent System for Payments), a four-level agent architecture that decomposes payment workflows into specialized roles: a Conversational Payment Agent handles user interaction, Supervisor agents manage task delegation, Routing agents direct requests appropriately, and a Process Summary agent validates completion. The hierarchical design allows both open-weight and proprietary LLMs to collaborate on payment automation, solving a previously unsolved problem in agentic systems.
Core Technical Contribution
The core novelty is the hierarchical decomposition of payment workflows into four distinct agent levels, each with specialized responsibilities, rather than attempting single-agent payment handling. This multi-level architecture explicitly separates concerns: conversational understanding, task supervision, intelligent routing, and outcome verification—allowing agents to operate within their competence boundaries. Prior agentic approaches treated payment as a generic workflow task, but HMASP recognizes that payment requires strict validation, security checks, and error handling that benefit from role specialization. The modularity also enables mixing open-weight and proprietary models, optimizing cost and capability across the agent stack.
How It Works
Input flows from user through the Conversational Payment Agent (CPA), which understands payment intent and parameters in natural language. The CPA then escalates to Supervisor agents at the second level, which understand the semantic structure of the payment task—recipient, amount, currency, deadline constraints. These supervisors delegate to Routing agents (third level) that make tactical decisions about payment method selection, risk assessment, and transaction routing to specific payment processors or APIs. Finally, the Process Summary agent validates that all steps completed successfully, reconciles the transaction state, and reports back to the user. Each level can independently query an LLM, allowing different model sizes/capabilities per tier and enabling retry/rollback at each stage.
Production Impact
For engineers deploying payment automation, HMASP eliminates the current choice between unreliable single-agent payment attempts and expensive human-in-the-loop workflows. You could integrate this by replacing manual payment verification steps with the four-tier agent pipeline, reducing payment processing time from hours to minutes while maintaining auditability. The modular architecture means you can start with small, cost-efficient models at the routing tier and scale to larger models only at the supervisor/summary levels where judgment matters most. Trade-offs include increased latency due to multi-hop agent calls (potentially 2-5 seconds overhead), higher LLM inference costs per transaction, and the need to instrument payment APIs for agent visibility. Integration complexity is moderate—you need structured payment processor APIs and state tracking between agent levels, but the separation of concerns simplifies testing and debugging individual tiers.
Limitations and When Not to Use This
The paper doesn't address adversarial scenarios—what happens when an LLM agent is prompted to transfer funds to the attacker or when payment APIs are compromised. It assumes payment processors expose well-structured APIs and don't discuss handling legacy or unstandardized banking systems common in emerging markets. There's no evaluation against fraud detection requirements or regulatory compliance (PCI-DSS, AML/KYC) that real payment systems require, and the abstract gives no metrics on accuracy, latency, or cost compared to baselines. The hierarchical design assumes clear task boundaries at each level, but real payment workflows often have exceptions and cross-cutting concerns (dispute handling, currency conversion edge cases) that may not decompose cleanly across four fixed tiers.
Research Context
This work builds on the emerging research in multi-agent LLM systems (expanding on frameworks like LangChain and AutoGPT) but applies them to the traditionally human-supervised domain of financial transactions. It follows the trajectory of recent papers on LLM-based workflow automation (Operator/Computer Use) while identifying that finance requires architectural innovations beyond generic agentic approaches. The hierarchical supervision pattern relates to prior work in multi-level reinforcement learning and human-AI collaboration, though here applied to financial domain constraints. This opens a research direction in domain-specific agent architectures, suggesting that other high-stakes domains (healthcare, legal) may benefit from similarly specialized agent hierarchies rather than general-purpose agentic frameworks.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::