---
title: "Nobody Wants Another Mainframe, Especially for AI - Blog - Coder"
description: "Developers switch models and coding agents every few weeks. Here's how platform teams keep up without rebuilding every time."
image: "https://www.datocms-assets.com/19109/1795050754-blog-hero-quiz-centered_2512x1348.png?fit=clip&fm=webp&w=800"
canonical: "https://coder.com/blog/nobody-wants-another-mainframe-especially-for-ai"
---

Nov 19 20266 min read

# Nobody Wants Another Mainframe, Especially for AI

[Matt Vollmer](https://coder.com/blog/author/matt-vollmer)

Share this article

Mainframes stuck around for decades because changing course cost too much. Nobody wants to repeat that with AI, and the teams responsible for rolling it out feel that pressure most.

Those Platform and AI Infrastructure teams often ask us some version of the same question: **“How can we quickly and responsibly get AI into our developers’ hands?”**

Some expect the answer to start with picking the right agent, but we don't have an off-the-shelf favorite. Our guidance depends on your requirements and what your developers _want_ to use. Cursor, Claude Code, Codex, or whatever ships next month could all be the right choice.

If you're a platform engineer, your job is to give developers what they need, even when that changes every other week. You should be able to say yes to the next agent without rebuilding everything underneath it.

We have exactly one opinion here at Coder: agents can only do serious work on infrastructure you control, and laptops or shadow VMs can't give you that at scale. Past that, how you run them is up to you, and Coder gives you several ways to do it from one deployment.

## What decides the path

We usually orient our guidance around how you answer [three questions](https://coder.com/agent-pathfinder) about your organization.

- Do the agent and the model it uses need to stay inside your network, or run fully offline?
- Do any of your developers want to keep using popular cloud-hosted agents?
- Do any of them want to work hands-on with an agent CLI in their terminal?

The first is about your requirements. The other two are about your developers’ preferences. Each one maps to one of the paths below.

![](https://www.datocms-assets.com/19109/1795050992-inline-three-paths-stack.png)

## Three paths, one Coder deployment

### When developers want to keep their cloud agents

Many developers started with cloud agents from Cursor and Claude Code because they're the easiest to get to and the friendliest to get running, especially for builders in developer-adjacent roles. Some enterprises can support them as they are. Others can't, because the agent runs and does its work in the provider's cloud, where your code leaves the network, and your internal systems are out of reach.

[Agent Relay](https://coder.com/docs/@main/agent-relay) changes where that work happens. Developers start an agent the same way they do today, and Coder builds a fresh, self-hosted workspace for each agent session. The agent's tool calls run in that workspace, on your infrastructure, and the workspace is deleted when the session ends. Your network policy decides what the agent can reach from the workspace.

The agent's orchestration and the model still run in the vendor's cloud, so code snippets and prompts still leave your network, and Coder can't see those AI requests or what they cost. For teams that are comfortable with that, it's the least disruptive option. For teams that aren't, Coder Agents, covered below, could be the better fit.

### When developers want an agent in their terminal

Developers who spend most of their day in a terminal often prefer an agent CLI like Claude Code or Codex, because it works right alongside their code, tools, and shell. On a laptop, that agent runs with the developer's own credentials and network access, which many security teams aren't willing to accept at scale.

Running those CLIs inside Coder workspaces keeps the same hands-on experience on your infrastructure. The agent can reach only what the workspace template allows, and ready-made agent modules in the [Coder Registry](https://registry.coder.com) make installation a few lines in that template. Those modules can also sign the agent in automatically. Without that, every fresh workspace means another login, which is the kind of friction that sends developers back to their laptops.

Model calls can go wherever you already send them. If you run your own model gateway, the agent can keep using it. For teams that want more than a proxy, [AI Gateway](https://coder.com/docs/@main/ai-governance/ai-gateway), part of Coder's [AI Governance](https://coder.com/docs/@main/ai-governance), ties every request to a developer and a workspace, records prompts, token usage, and tool calls, and [stops a developer's requests once they reach their budget](https://coder.com/docs/@main/ai-governance/ai-gateway/cost-controls). Whether this path can run fully offline depends on whether the agent supports a self-hosted model.

### When agents and models need to stay inside your network

Some teams can't send code or prompts outside their network at all, and some work in environments with no internet access. For them, the agent loop, the model, and the execution all have to live on infrastructure they control. That's the case [Coder Agents](https://coder.com/docs/@main/coder-agents) was built for.

Its agent harness runs on your Coder control plane, calls only the model providers you configure, including self-hosted models or models served through AWS Bedrock, and does the work in a Coder workspace. Provider credentials stay in the control plane and never enter the workspace. With a self-hosted model, the whole system can run air-gapped. Developers hand off tasks through a chat UI or the API, which also makes it useful for work kicked off from CI or other automation.

It's worth saying that running your own models isn't automatically cheaper. GPUs are expensive, and the teams that choose this path usually do it because they need the control. And different [organizations](https://coder.com/docs/@main/coder-agents/platform-controls/organizations) in the same Coder deployment can work under different levels of control, with some using only self-hosted models and others using models from commercial AI providers.

## What it costs to change course

Many of our enterprise customers support several of these deployment models on a single Coder deployment. Coder Agents for teams that can't let code leave the network, Agent Relay for the Cursor and Claude Code fans, and agent CLIs for developers who are most comfortable in a terminal.

Different people work on different projects, in different parts of the codebase, under different data classifications, and those differences call for different ways of running agents. Supporting each one shouldn't mean re-architecting or standing up separate infrastructure. Coder gives you that choice on one platform, so when the next business requirement comes down or the next great agent ships, the cost of switching drops out of the decision. You already have what you need to support it.

## Find your paths

[Agent Pathfinder](https://coder.com/agent-pathfinder) is part of [Run Week](https://coder.com/run-week). Answer three questions for your whole organization and it shows which paths fit, plus what each one means for credentials, offline teams, and keeping a record of AI use. "Not sure yet" is a fine place to start.

Want to talk it through? [Get a demo](https://coder.com/demo). Or [install open source Coder](https://coder.com/docs/@main/install) and try it yourself. And if your teams have a constraint the guide doesn't cover, tell us. Whatever your developers want next, bring it.

[![Matt Vollmer](https://coder.com/_next/image?url=https%3A%2F%2Fwww.datocms-assets.com%2F19109%2F1778785451-1767660038155.jpg%3Ffit%3Dcrop%26fm%3Djpg%26h%3D100%26w%3D100&w=2048&q=75)](https://coder.com/blog/author/matt-vollmer)
[Matt Vollmer](https://coder.com/blog/author/matt-vollmer)

Product Marketing Manager at Coder

Matt is a product marketing manager specializing in developer tooling. With a background in design, development, and product management, he has worked in high-growth startups, helping teams build and scale software for today’s engineering challenges.

[Learn more about Matt Vollmer](https://coder.com/blog/author/matt-vollmer)

### Subscribe to our newsletter

Want to stay up to date on all things Coder? Subscribe to our monthly newsletter for the latest articles, workshops, events, and announcements.
