AI Environments

We typically work with AI within an environment. Most “traditional” tools are now flavored with AI suggestions, including messaging and email apps (which can be optionally turned off). There are now innumerable AI tools for coding and analysis, and this will not be a compendium. Instead, I am trying to keep abreast of the landscape that seems relevant to me, and will add notes here as I learn more.

What is an AI Environment?

An AI environment is a setting where artificial intelligence systems operate and interact with data and users. It includes hardware, software, data, and algorithms that facilitate AI functions. Environments can be physical (like robots) or virtual (like simulations or cloud platforms). AI environments allow for real-time interaction and feedback between AI systems and their surroundings. They often incorporate machine learning capabilities, enabling AI to adapt and improve over time. Common applications include autonomous vehicles, smart homes, and virtual assistants. (Google Gemini)

Integrated Development Environments (IDEs)

Initially, it seemed important to understand Model context protocol (MCP), the concept and protocol developed by Anthropic (and first available in Claude Code). The revolutionary idea is that a large language model (LLM) is not just a text generator, but can interact with your other tools and data to get information and take actions. However, this was quickly integrated into an integrated development environment (IDE) building on Visual Studio Code (VScode or code) so that one can edit code manually and/or interact with an LLM via an AI agent in a side panel. This AI agent uses MCP to connect with your data and code under your guidance. We need not be concerned about the MCP mechanics of LLM communication with local (or cloud-based) tools and data. Instead, we carry on conversations with an AI agent who, with explicit permission from us, changes local code or documents and creates reports.

Cloud-Based Research Environments

Cloud-based research environments enable researchers, including students and instructors, to develop code, optionally with AI agent interactions, without having to try to set up systems on their own, idiosyncratic devices.

For those working directly in GitHub, code generation within repos is improving with the nuanced integration of GitHub Copilot, both for repo editing and for use in GitHub Codespaces, which is a cloud-based IDE optionally integrated with each GitHub repo. GitHub also has capabilities to publish GitHub Pages, such a personal or organization portfolios.

Google Colab provides a free and simple way to code in R and Python, with optional AI support via Google Gemini.

CyVerse Discovery Environment is a cloud-based enviroment for data science and bioinformatics. I have used this in conjunction with Environmental Science and Innovation & Impact Laboratory (ESIIL) to catalyze instruction, research and tool development. Recently, CyVerse DE added AI engines Roo and Cline to their VScode IDE environment. (NB: Roo and Cline seem to work best in a Chrome browser.) We can use LLM APIs from AI Verde to tunnel a variety of LLMs into either CyVerse DE or other IDEs of our choice. See AI-VERDE Manual & Resource Site and AI Verde API.

CyVerse and AI Verde use Jetstream2 to host computing tools at scale. Jetstream2 is part of the NSF-funded ACCESS, which provides free access to compute and AI resources at scale. These resources also include HTCondor and PATh. ACCESS resources typically require an educational account and benefit from technical guidance.

Google Gemini and Antigravity

Google Gemini has an enterprise contract with UW-Madison; hence I concentrate a bit on it here. Much of this applies more broadly to other AI environments. While Gemini is available via the browser, it is also being embedded into a variety of tools. These tools are evolving fast, due to competition as well as tool advances through use of AI on the tool development process.

Google initially offered Antigravity 1.0 as an extension of VScode with a Gemini AI agent in the right panel. On 19 May 2026, Google released Antigravity 2.0 as a suite of 4 tools:

  • Antigravity 2.0: full multi-agentic AI code development environment
  • Antigravity CLI: command line interface (CLI), used in the terminal
  • Antigravity SDK: software development kit (SDK), used by developers to build apps
  • Antigravity IDE: integrated development environment (IDE), an updated version of the previous (1.x) IDE

Unfortunately, this release confused previous users. The new tool, with the same name as the previous tool, presents as a chat window without IDE-like features, and without clear indication about how it differs from the prior tool, or how to get back to the prior tool. This will be a useful tool for vibe coding but not directly relevant to those of us developing code-based tools. Antigravity IDE seems better suited for that task. Again, unfortunately, the IDE did not carry any previous chat history (known as conversations) over to 2.0. See Recovering Antigravity Conversations for a guide I wrote to recover them based on community response.

Other AI Environments

Antigravity and many other IDEs began as forks of the VS Code that integrate LLMs with collaboration on a user’s local files and tools. Some have evolved away from these roots toward standalone apps.

See caution in Using R in VS Code with Radian about radian and AI environments.

Agentic AI

There are concepts of “AI-native” and “agent-native” IDEs, but I am not sure I fully grasp the distinction. It seems that “AI-native” concerns conversations with LLMs, while “agent-native” concerns the use of AI agents to do work. Further, the straightforward use of conversations via prompt and context engineering is often described (as below) in the setting of using one AI agent at a time.

Nowadays (Spring 2026), sophisticated users of AI orchestrate multiple AI agents to accomplish complex tasks. Claude, [OpenAI ChatGPT][https://openai.com/index/introducing-workspace-agents-in-chatgpt/) and Google Antigravity have been evolving such capabilities quickly in the commercial world.

SOUL.md

There is an emerging concept of SOUL.md that defines each agent by the content of a markdown file in a project folder.

This all takes time and practice to build up both our ability to understand and use multiple agents. Selection and organization of appropriate agents for a task. Further, multiple agents will likely require a larger monthly fee, if using commercial agents.

Agentic AI References

What is an AI Harness?

According to Andrew Maynard, “the crystallizing moment came in early February 2026, when Mitchell Hashimoto — cofounder of HashiCorp and creator of Terraform — published a blog post that gave the … name … “harness engineering”…. “It is the idea that anytime you find an agent makes a mistake, you take the time to engineer a solution such that the agent never makes that mistake again” (Hashimoto, 2026). … And on February 18, Ethan Mollick’s widely read guide to AI both popularized and started the process of normalizing the term as it organized its entire framework around three concepts: “Models, Apps, and Harnesses” (Mollick, 2026).”

As stated by Jaymin West, “Ethan Mollick introduced the horse harness metaphor to explain why the term is well-chosen (Mollick, ~2026-Q1). A horse has raw physical capability — strength, speed, endurance. But raw capability cannot pull a plow, draw a carriage, or haul freight without a harness. The harness converts raw power into directed, controlled, useful work.” Sarah Chen stated, “The Large Language Model (LLM) is a powerful horse, enormous raw capability, but no sense of direction, no understanding of boundaries, and no concept of “stop.” The harness is the bridle, reins, and saddle. It channels that power into controlled, useful work. Without it, the horse runs wherever it wants.”

Google Gemini attributed the horse metaphor for AI harness to the late John McCarthy, founder of the field of AI, although I have not been able to find the term connected with his name. Interestingly, there is a horse photographer named Jon McCarthy, and a (deceased) noted breeder of Irish Draft Horses named John McCarthy Perhaps Gemini conflated these?

Here are some other analogies:

  • “The model is commodity. The harness is moat.” (Sarah Chen)
  • “The model is the engine. The harness is the car.” (Grant Harvey)

According to Google Gemini, An AI harness is a framework designed to integrate AI capabilities into applications. It provides tools and resources for developers to build and deploy AI models efficiently. The harness often includes pre-built algorithms and data processing pipelines. It facilitates the management of AI workflows, from training to deployment. AI harnesses can enhance collaboration among data scientists and engineers. They help streamline the process of scaling AI solutions across different environments.

Harness Engineering References

Table of contents

This site uses Just the Docs, a documentation theme for Jekyll.