Mainframe Mediators: AI Agents' Growing Pains

The Current State of Play

AI agents have made significant strides in automating tasks, including those performed by credit controllers [1]. These intelligent assistants can handle various responsibilities such as invoicing, monthly billing runs, payment bank verifications for fraud prevention, and expense verification and proof validation. However, AI agents are not without their limitations. One major challenge is ensuring interoperability with mainframe systems, which often require specific protocols and interfaces to communicate effectively [2].

Open-Source Solutions and Proprietary Proposals

When it comes to AI agent development, there are two primary approaches: open-source solutions and proprietary proposals.

OpenCode [1] is an open-source agent that enables developers to write code in their terminal, IDE, or desktop. Its features include automatically loading the right LSPs for the LLM, multi-session support, shareable links, GitHub Copilot integration, and support for local models through Models.dev [2]. OpenCode's architecture allows for seamless integration with a wide range of programming languages, making it an attractive option for developers looking to build custom AI-powered applications.

On the other hand, Hopper [2] is a proprietary AI agent that provides separate builds for Apple Silicon and Intel processors on macOS. It also offers an x64 installer on Windows and supports Debian and Ubuntu via a.deb package and any Linux distribution via an AppImage. Hopper's focus on compatibility with popular operating systems makes it an excellent choice for developers working on diverse projects.

Feature	OpenCode	Hopper
LSP support	√
Multi-session	√	√
Shareable links	√
GitHub Copilot integration	√

Mainframe Interoperability: TN3270 and FTP

Hopper [2] connects to z/OS using two standard mainframe protocols: TN3270 for the live terminal session and FTP for datasets and the JES queue. This allows developers to read COBOL source, submit and monitor batch jobs, decode VSAM records, and write changes under user supervision.

To illustrate the benefits of Hopper's connection methods, consider a scenario where a developer needs to access a mainframe system to retrieve data from a VSAM record. Using Hopper, they can establish a secure connection via TN3270 and FTP, allowing them to read and manipulate the VSAM records with ease.

Separation of Concerns: Read-Only Actions vs. Mutating Actions

Hopper's approval model [2] separates read-only actions from mutating actions, requiring explicit user approval for the latter. This ensures that AI agents can perform tasks without unintended mainframe changes.

In practice, this means that when a developer wants to write data to a VSAM record using Hopper, they must explicitly approve the action before it is executed. This added layer of control helps prevent accidental changes to critical system data.

Action	Hopper Approval Model
Read-only (e.g., listing members)	Instant execution
Mutating (e.g., writing a member)	Requires explicit user approval

Illustration: Comparison of OpenCode and Hopper Features

Performance Considerations

While AI agents like OpenCode and Hopper can significantly streamline tasks, they are not without their performance implications. One key consideration is the potential for increased CPU usage, particularly when handling complex workflows or large datasets.

For instance, a developer using OpenCode to automate a financial reporting task may notice a slight increase in CPU utilization during peak processing times. However, this overhead can be mitigated by optimizing code, leveraging multi-threading, or utilizing cloud-based services with scalable infrastructure.

Conclusion

In conclusion, the current state of AI agents for mainframe systems is marked by significant advancements and challenges. Open-source solutions like OpenCode offer flexibility and customizability, while proprietary options like Hopper provide compatibility and ease of use. As developers continue to push the boundaries of what is possible with AI-powered mainframe automation, it will be essential to balance the benefits of these tools against the potential performance implications.

[1] OpenCode. (n.d.). Retrieved from https://opencode.ai/

[2] Hopper. (n.d.). Retrieved from https://www.hypercubic.ai/hopper

[2] Hopper. (n.d.). Retrieved from https://www.hypercubic.ai/hopper

[2] Hopper. (n.d.). Retrieved from https://www.hypercubic.ai/hopper

[3] The system allows specialists to focus on care, reduce administrative burden, and ensure faster, more reliable insurance payments. from: https://news.ycombinator.com/item?id=43280836 supporting quote: This allows specialists to focus on care, reduce administrative burden, and ensure faster, more reliable insurance payments.

[3] Cenote can function in a headless manner, without requiring user interaction. from: https://news.ycombinator.com/item?id=43280836 supporting quote: By this, we mean that all core functionality—data extraction, EHR integration, and even back-and-forth communication with referring providers—does not explicitly require a UI for user interaction.

References

1. The open source AI coding agent — opencode.ai (20 March 2026)
2. Agentic TN3270 Emulator With Hopper | Hypercubic — hypercubic.ai (12 May 2026)
3. Launch HN: Cenote (YC W25) – Back Office Automation for Medical Clinics — news.ycombinator.com (6 March 2025)