Cortex-FL: Agent-Mediated Collaboration for Federated Learning

Rui Sun, Jinfeng Peng, Zehua Cheng, Jiahao Sun, Zhipeng WangarXiv 2026 · arXiv

TL;DR

Cortex-FL is an "agent-native" federated learning (FL) system that allows users to coordinate complex, multi-party model training using simple natural language instructions. By delegating the "logistical nightmare" of finding collaborators, negotiating terms, and managing blockchain-based payments to AI agents, it lowers the barrier for organizations to participate in privacy-preserving collaborative AI.

What Problem Does This Paper Address?

Federated Learning (FL) is a powerful paradigm that allows different organizations to train a shared AI model without ever exchanging their raw, sensitive data. However, in practice, setting up a "cross-silo" federation is incredibly difficult.

Currently, an organization wanting to start an FL project must manually identify trustworthy partners, align on technical objectives, configure compatible execution environments, and establish complex incentive structures. This requires significant specialized expertise and manual coordination. As shown in Figure 1, the authors contrast the current manual "hand-coordinated" state of FL with their vision of an integrated, agent-mediated workflow.

Figure 1: From (a) agentic service delegation and (b) manually coordinated FL to (c) agent-mediated FL in Cortex-FL.

Methods

Cortex-FL introduces Agentic Nodes (AN). Each data owner operates a node where an AI agent acts as their representative. The system architecture is built on several key pillars:

  1. Natural Language Delegation: A user can simply type, "I need a 5-class model for cats, dogs, birds, ships, and trucks." The agent then translates this into a technical task specification.
  2. The Trust Model: Crucially, agents coordinate but do not hold ultimate authority. The system uses "trust gates"—specific points in the workflow (like spending money or accessing a dataset) where the human owner must provide explicit approval.
  3. Integrated Lifecycle: The system automates the entire FL lifecycle:
  4. Discovery: Using AgentaNet to find peers with relevant data.
  5. Negotiation: Establishing bilateral agreements and service fees.
  6. Execution: Running training in sandboxed environments using FL Alliance and FLocKit.
  7. Settlement: Using the x402 protocol on the Base Sepolia blockchain testnet to handle payments in USDC once the model is verified.

The full architecture and the six-stage task lifecycle are detailed in Figure 2.

Figure 2: Architecture (top) and task lifecycle (bottom) of Cortex-FL.

Datasets

The authors demonstrated the system using a five-class image classification task based on the CIFAR-10 dataset.

  • The Scenario: To simulate a real-world need for collaboration, the "organizer" node only held data for two classes (cats and dogs), while three "peer" nodes held the data for the remaining classes (birds, ships, and trucks). This forced the organizer to find and negotiate with peers to achieve a functional 5-class model.

Key Insights and Ideas

  • Separation of Coordination and Authority: This is the "aha moment" of the paper. By letting agents handle the tedious planning and communication while keeping the human "in the loop" for critical authorizations, the system remains secure without being a burden to manage.
  • Incentives as a First-Class Citizen: Most FL research focuses on the math of training. Cortex-FL treats the economy of the training as equally important, integrating automated service fee negotiation and blockchain settlement directly into the training loop.
  • Verifiable Delivery: The system doesn't just pay peers for participating; it uses a "signed artifact chain" to ensure that the model delivered matches the task originally agreed upon before releasing funds.

Key Results

The authors successfully deployed Cortex-FL across a four-node setup (a MacBook Pro and three Mac minis).

  • End-to-End Automation: A model requested in natural language successfully moved through discovery, contract negotiation, training, and payment.
  • Performance: The collaborative training successfully reduced validation loss from 1.52 to 0.96 over 10 rounds.
  • Economic Settlement: The system successfully settled service fees (0.006 test USDC) across the nodes upon verified delivery of the model.

The user interface, shown in Figure 3, provides a comprehensive "task report" that tracks everything from the roster of peers to the real-time validation loss and blockchain activity.

Figure 3: The Cortex-FL workspace, with the task report (roster, round timeline, validation loss, payments, and on- chain activity) on the left and the agent chat on the right.

Limitations and Future Work

While a successful proof-of-concept, the authors note several areas for growth:

  • Simple Aggregation: The current implementation uses equal-weight parameter averaging; more sophisticated aggregation methods could be integrated.
  • Template Dependency: The system currently relies on pre-defined model templates (e.g., for image classification or LLM fine-tuning).
  • Scalability: The demonstration was limited to four nodes; future work will need to explore how agent negotiation scales as the number of potential collaborators grows into the hundreds or thousands.

Summary updated

Abstract

Federated learning (FL) enables independent data owners to train models collaboratively without centralizing raw data. However, identifying collaborators, preparing training tasks, and coordinating execution and incentives hinder its practical adoption by organizations without specialized expertise. We present Cortex-FL, an agent-native FL system that enables users to express learning objectives in natural language and delegate the collaboration required to train a model. Agents translate these objectives into shared tasks, discover collaborators, and negotiate participation agreements, connecting collaboration formation with distributed training and incentive management. The architecture separates agent coordination from local execution authority, allowing owners to delegate coordination while retaining control over their data, computation, and financial commitments. We demonstrate the complete workflow on a four-node deployment, in which a model requested in natural language proceeds through collaboration formation, contract-governed training, verified model delivery, and x402 settlement on a blockchain testnet.
BibTeX
@misc{sun2026cortex,
  author = {Rui Sun and Jinfeng Peng and Zehua Cheng and Jiahao Sun and Zhipeng Wang},
  title = {{Cortex-FL: Agent-Mediated Collaboration for Federated Learning}},
  year = {2026},
  eprint = {2611.17186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17186},
}
Figure