AIOS names a kernel and a personal app. A company needs an open-source layer it owns.
The name AIOS covers an agent kernel and a personal app; Kortix is the open-source CompanyOS a company keeps in its own repo.
agiresearch's AIOS gives agents a kernel that schedules their requests against the model backend it is configured with, and Kortix is the open-source CompanyOS that keeps a company's agents, memory and configuration in a repository it owns. The same four letters also name The AIOS, a personal app. A third meaning comes from an implementation partner: a layer that sits over a company's systems. Which of the three is the layer a company with shared systems actually buys?
Two products and a definition share four letters
The four letters usually expand to AI Operating System. Two products carry the name. agiresearch's AIOS calls itself the AI Agent Operating System and ships as a kernel plus an SDK. The AIOS is a personal system that turns one person's AI use into a team of co-workers. Innoworks, a consultancy that implements AI systems, gives the third meaning. That layer sits on top of the CRM, the administration system, the document management system and the customer portal. It reads from all four at once and runs processes at the organizational level.
What each one manages decides which of them a company can use as its layer. A kernel manages resources for agents. A personal app manages one person's context and a small team's shared space. The company layer reads the systems the company already runs and changes records inside them. Neither product is that layer.
Kortix does not put those four letters on the layer either. The category it builds is CompanyOS, and the company keeps the whole of it.
A kernel schedules agents
agiresearch's AIOS is a research project, and the thing it schedules is the agent. The kernel holds the model, memory, storage and tool resources an agent needs. It keeps scheduling, context, memory, storage and access control out of the agent application. The paper reports up to 2.1x faster execution for agents built with other frameworks.
Running it takes a Python 3.10 or 3.11 environment and a configuration file. Provider keys go in aios/config/config.yaml. The docs set out five deployment modes, two of which are implemented. The kernel hands results back through the SDK to the application that called it. The latest release, v0.3.0, dates from January 2026.
A personal app that starts with one person
The AIOS is built for one person first and opens out to a team. Its own line is "AI as a team, not a tool", and the site calls the design "multiplayer by architecture". A teammate runs --mount {repo-url} and inherits the venture's brain. A shared space keeps its files on Drive, GitHub or a local folder. The framework is plain markdown, and the vault it scaffolds is a separate private space on your own machine. Nothing runs on their servers: "There is no AIOS server. Nothing phones home, nothing is collected, no account is created."
Work that has to continue while the owner is away runs on a "two-machine architecture". A primary MacBook does the day-to-day work, and an always-on Mac mini hosts the autonomous fleet, walled off at the operating-system level. INTENT.md is the trust contract, and it records what may run without the person. The vault syncs only where its owner chooses. Finished work arrives as an edit in that vault, which suits a person or a small team whose reviewer is the owner.
The decision is ownership
A company is choosing an owner for four things. They are the agents, the memory, the configuration behind both, and the route a finished piece of work takes to production. What each agent may do, what it remembers and where the record of its work lands all follow from that choice. Each of the three meanings answers it differently.
| What the company decides | Kortix | agiresearch's AIOS | The AIOS |
|---|---|---|---|
| Open source | Yes | Source is public; no license published | Yes |
| Where configuration lives | kortix.yaml and markdown in the project repo | aios/config/config.yaml on the host | Markdown and INTENT.md in a private vault |
| Where memory lives | memory/ in the project repo | The kernel's memory manager | A private vault on the owner's machines |
| How work lands | A change request toward main; merge is default-deny for agents | Back to the calling agent application | As edits in a private vault |
| Compute boundary | One isolated sandbox per session | The machine running the kernel | The owner's Mac, or a second Mac |
The kernel runs where someone installs it, and the personal system runs on the owner's machines. A company layer has to give each piece of work its own machine and land the result where other people can review it.
What a company layer has to hold
Kortix starts from the repository. The project is the repo, kortix.yaml at its root is the governance, and the agent's behavior sits beside it in markdown. The manifest decides which connectors, secrets, skills and project permissions a given agent gets. A grant the file leaves out is one the agent does not have. A memory/ folder carries the project's purpose, its connections, its conventions and its decisions across sessions. MEMORY.md lists the files that hold them.
default_agent: kortix
agents:
kortix:
connectors: all
secrets: all
skills: all
kortix_permissions: allThat block is one agent's reach, and deleting a line takes that category away. A trigger in the same file starts a session with the agent it names, in a fresh session by default, with no person present. When an agent edits the memory/ files, the change lands as a diff on its branch and reaches the rest of the project only through the merge.
Session work reaches main through a change request, and merge is default-deny for agents. An agent merges only when its kortix_permissions grant includes project.gitops.merge. Each session gets its own isolated sandbox, so one agent's mess cannot reach the next piece of work.
Open source. Read it, fork it, audit it. The longer case for treating the company itself as the repository is in the post on owning the layer.
When another answer is the right one
A research group studying how a kernel schedules agent requests is better served by agiresearch's AIOS. There the kernel is the subject, and a review step adds nothing to the study. A person or a small team whose work lives in a vault is better served by The AIOS. There the file is the deliverable and the owner signs off. A company is the case left over. Its work touches shared systems, several people have to agree before a record changes, and a security reviewer will ask to see the configuration.
The layer a reviewer can read
The four letters name two products and a definition, and the definition is the one a company has to fill. A kernel serves agents and a personal app serves a person. A company needs the layer that keeps the agent's definition, its memory and its permissions in something the company owns. That is what Kortix is, open source, in a repository a company can read, review and revert.