The Agentic OS Is NOT

It's not an operating system.

There’s been a lot of noise about “the Agentic OS”, an operating system that will integrate AI and make it seamless.

Microsoft is doing a quintessential Microsoft, trying to mix the new thing with their bloated, poorly conceived mess, seeking to control it. They might dominate the desktop, but they do so from the position of the least (and most) common denominator. And I’m using the word common in its old fashion sense - vulgar.

The organization doing things the “Unix way” is Anthropic. There’s a reason they’ve captured the corporate development environment. What I see emerging in ~/.claude is the same thing one sees on Unix operating systems - files and folders with clear purposes, easily debugged, modified, and restored if there’s trouble.

Let’s delve deeper into this.

Attention Conservation Notice:

Pretty technical screed about why Anthropic is winning by not doing anything complex, controlling, invasive, or stupid. If that black terminal window scares you, probably best you scroll on to safety.

Observations:

This is what’s in my ~/.claude folder today - I gave Claude the directory listing and it gave me this old school ascii graphic of the contents. Looks like something you’d find in a 1980s Unix oriented zine, doesn’t it?

The plugins folder is the package manager. The skills are the system packages, stuff that runs in the background. The agents are for both user and system. The commands are akin to the aliases in your shell’s profile. Those of us who are Unix literate will make a copy of this stuff just in case, then wade in swinging. It’s familiar enough we feel confident laying hands on it.

|-- EXTENSIBILITY PRIMITIVES (the good stuff)

| |

| |-- agents/ Subagents - parallel workers with specialized roles

| |-- commands/ Slash commands - user-invoked shortcuts (/command)

| |-- skills/ Model-invoked capabilities - Claude decides when to use

| |-- plugins/ Distribution packages - bundles of the above

|

|-- CONFIGURATION

| |

| |-- settings.json User preferences and feature toggles

|

|-- RUNTIME STATE (Claude’s working memory)

| |

| |-- history.jsonl Conversation history log

| |-- file-history/ Recent file operations

| |-- projects/ Project-specific state and context

| |-- plans/ Multi-step task planning artifacts

| |-- todos/ Task tracking

| |-- session-env/ Environment variables per session

|

|-- INTERNALS (here be dragons)

| |

| |-- debug/ Diagnostic logs

| |-- ide/ IDE integration state

| |-- shell-snapshots/ Terminal state captures

| |-- statsig/ Feature flag / A-B testing data

INVOCATION MODEL:

commands/ --> YOU type /command to trigger.

agents/ --> YOU or CLAUDE can invoke.

skills/ --> CLAUDE decides autonomously based on task context.

plugins/ --> Installed packages that provide all of the above.

Containerization:

We’ve been using varying degrees of virtualization since CPUs sprouted paged memory management, now many decades in the rearview mirror. Today I see a variety of containers in use in my environment.

• Claude Desktop has a “sandbox” Linux system for code execution.

• OrbStack replaced Docker on my Mac, mostly for Postgres.

• VMware offers a whole Ubuntu Linux for certain high risk environments.

• Proxmox on machines behind me runs Linux VMs.

• Those Linux VMs run Docker.

• Except for the ones involved in Microk8s

So there’s one thing that provides a full OS to Claude’s agent behavior, and a bunch of stuff that provides APIs and MCP servers that are accessible by the three clients (Antigravity, Claude Desktop, Perplexity) I am using.

Containers are compartments, that’s always been true, it’s not just maintenance, it’s for partitioning job duties. If you noticed Fleeing SQLite3, Finding Postgres you already know I changed databases specifically to put an end to Claude’s tableshitting. An LLM on the loose is a goat rodeo - they have to be put into a harness and made to do what you want, which is much more focused than the sum of what you tell them.

There are some “Agentic OS” offerings out there, I looked at what Perplexity found, there are varying levels of maturity, varying directions. None of it excites me to the point that I will mentioned it in our engineering channel in hopes someone else will go exploring.

Conclusion:

As of today, my AI work surface is a triumvirate:

• Antigravity - an IDE using Claude Opus4.5 and Gemini3/ChatGPT5.

• Claude Desktop/Code - the nexus for both coding and agent innovation.

• Perplexity - A qualitative, search oriented system replacing my ChatGPT.

If you read Choosing Your AI Model(s) you know that the big three - Gemini3, Opus4.5, and ChatGPT5 have differing strengths and weaknesses. Antigravity and Perplexity permit you to alternate among the three, King Claude stands alone.

Looking at it another way:

• Claude - creating the agentic foundation.

• Antigravity - put a graphical interface on the foundation.

• Perplexity - best features for my sense making needs.

Antigravity leapfrogged Cursor/Windsurf, it’s Google’s move to displace Anthropic from its king of the hill status with the enterprise. I will always have a development GUI with these features, but there may well be an Anthropic fork of VSCode next quarter that displaces Antigravity.

Perplexity is unique in what it does, I’ve known it featured RAG (Retrieval Augmented Generation), but its search first, then respond, combined with “always provide citations” is a relief in the wasteland of hallucinated slop. The availability of citations mirrors what I have done with Parabeagle.