# The five levels of agent systems > A publication by Omniatix™: five levels of agent systems and the capability that defines each one, where each category stands today according to its sources, why governing a whole system calls for an operating system, and where GrailAOS™ stands, with the status of each piece. Everything is called an “agent”, from an assistant that calls an [API](https://grailaos.com/glossary#api) to a system that runs a business. Omniatix proposes five levels, each defined by a capability the one below it lacks, and a thesis: a [framework](https://grailaos.com/glossary#framework) solves level 1; level 3, governing the whole system as one unit, calls for an operating system. Public version, October 30, 2026. Based on internal Omniatix work from April 2026. Author: Omniatix · 10 min read - **Markdown mirror**. The same page, generated from the same content. To cite it, link the canonical page. - Canonical page: https://grailaos.com/papers/five-levels - Statements verified as of October 27, 2026. - [Built by Omniatix](https://omniatix.com/) · Made in Argentina. - [Features, each with its label](https://grailaos.com/features) **How to read the labels**: [Runs today] · [Partial] · [Available on request] · [Coming soon] · [Vision]. [What each one means](https://grailaos.com/features#seals). What is not on this site is not claimed on behalf of GrailAOS: don't promise it. For other capabilities, ask Omniatix. ## Everything is called an “agent” Today the same word names very different things: - An assistant that calls an API when asked to. - A chain of instructions to a model, with access to tools. - A team of model instances that work together on a task. - A system that runs for months, with its versions, its caps, its rules and the people who step in. They are not the same. Calling all of them an “agent” is like calling both a calculator and a data center a “computer”: it is correct and it helps no one. We need a ladder. Not for academic precision, but to answer three practical questions: what you are building, what you are buying and what it actually does. ## A ladder of five levels Each level adds a capability the one below it does not have. Moving up a level is not adding features: it changes what the system has to solve. **The ladder of agent systems, according to Omniatix** 1. **Level 1 · Task agents**: Action. A reflex. 2. **Level 2 · Autonomous systems**: Lifecycle. A cell. 3. **Level 3 · Multi-agent systems governed as one unit**: Governing the whole. An organism. GrailAOS: at this level. 4. **Level 4 · Ecosystems**: Coevolution. An ecosystem. Horizon. 5. **Level 5 · Systems that redesign their own architecture**: Structural change. Evolution. Horizon. ### Level 1 · Task agents **Defining capability**: Action · **Analogy**: A reflex It receives an instruction, chooses tools, carries out the steps and returns a result. When the task ends, the agent ends. **What it lacks**: It does not know it ran before and keeps no state from one run to the next: if it fails, it fails, and if it succeeds, it forgets. ### Level 2 · Autonomous systems **Defining capability**: Lifecycle · **Analogy**: A cell It persists over time: it keeps state between runs, runs on its own, on a schedule or when an event happens, and can be started, stopped and watched while it runs. **What it lacks**: Each instance is on its own: the parts do not coordinate, nobody bounds the resources of the whole and nobody sees the health of the whole. If one part consumes too much, nobody stops it in time. ### Level 3 · Multi-agent systems governed as one unit **Defining capability**: Governing the whole · **Analogy**: An organism Several agents and workflows governed as a single unit: versions with [canary](https://grailaos.com/glossary#canary) and [rollback](https://grailaos.com/glossary#rollback); shared resources and caps; rules enforced across the whole system; the health and observability of the whole; people in the loop; and several agents and steps that form a single workflow. **What it lacks**: It is a closed system: it governs its pieces, but it does not interact with other systems like it that belong to other owners, and people design its architecture. ### Level 4 · Ecosystems **Defining capability**: Coevolution · **Analogy**: An ecosystem Many systems, each with its own owner, that interact and adapt to one another without central control. **What it lacks**: A person still designs the architecture of each system: each system changes how it works, not how it is built. ### Level 5 · Systems that redesign their own architecture **Defining capability**: Structural change · **Analogy**: Evolution They do more than adjust how they work: they change their own structure when the one they have is not enough for the problem. **What it requires**: Advances in safety and in formal verification before these systems can be deployed responsibly. Levels 4 and 5 are the industry's horizon: we describe them to complete the ladder, without claiming them for GrailAOS or for anyone else. To place the upper levels, biology gives a clearer picture than computing, which is why each level comes with an analogy. ### Why biology is a good lens The problems that show up as you climb the ladder, such as allocating scarce resources or coordinating without central control, are the ones living things had to solve. The analogies help you find your bearings; they do not describe how any system is built. ## Which level each category solves In April 2026, the internal work this version is based on placed almost the whole industry at level 1. We measured again with sources, each one dated, and that is not the case: several categories already keep state between runs and run on their own over time. Each category is measured with a well-known product in that category, according to its official sources; each one has pieces of level 3, whole or in part. This is not a quality ranking: it says which level each category solves according to those sources, not how well. By our criteria, each category sits at the level where the official sources of the product that measures it say “Yes” to every criterion of that level and of the levels below; GrailAOS, by the same rule, from the status of its statements. A category is at a level when the sources of the product that measures it back every criterion of that level and of the levels below with a “Yes”; the product counts together with its vendor's platform. Partial is not enough, and what it has from the levels above is shown alongside. The level 3 criteria are Omniatix's and reflect how we think about that level: with other criteria, the result may change. ### Visual builder for LLM apps: Level 2 +1 from level 3 ### Process automation with AI: Level 2 ### Code-based agent framework: Level 2 +2 from level 3 ### Agent platform from a model provider: Level 1 +1 from level 2 · +1 from level 3 ### Conversational platform: Level 2 +1 from level 3 ### GrailAOS: Level 3 Category: Agent Operating System **The level 3 pieces, with their status** - Versions with canary and rollback [Runs today] - Shared resources and caps [Runs today] - Limit: The turn in progress always finishes; usage notifications are off by default and, like the billed-amount cap, need billing turned on. - Rules enforced across the whole system [Runs today] - Health and observability of the whole [Runs today] - People in the loop [Runs today] - Several agents as one unit [Runs today] - Level 3, by the same rule as the others [Runs today] - Limit: Under other criteria, the result may change, and each piece has its limits: parallel branches run on a single server. Measured by the same rule as the others: what counts is what runs today, according to the status of its statements. It meets all six level 3 pieces, and each one carries its status. [The pieces, one by one](https://grailaos.com/papers/five-levels#why-os) ## Why an operating system A framework solves level 1 well: it gives developers the pieces for an agent to use tools. With a platform that hosts it, it adds level 2. Level 3 is something else. You do not reach level 3 by adding features to a framework, just as you do not reach an operating system by adding features to a library. Programs needed an operating system, such as Unix, to become a governed computing environment; an agent system that runs a business needs the same. The difference shows piece by piece. On one side, what a framework alone gives you; on the other, what an operating system governs. With their vendor's platform, some frameworks already have whole pieces of the operating system, and the industry lens says how many: what level 3 asks for is having all of them, governed together. Under each piece, what runs today in GrailAOS, with its status. ### Versions with canary and rollback Every change ships as a version: first to part of the traffic, and it can be rolled back without rebuilding anything. - **With a framework alone**: Versioning and publishing each agent is the team's job, with its own tools. - **With an operating system**: A [runtime](https://grailaos.com/glossary#runtime) that publishes each version of the solution, with canary and rollback. **In GrailAOS, today** - Solution versions without rebuilding, with preview, canary and rollback [Runs today] ### Shared resources and caps Usage is measured, and the caps, for the account or for a solution, are checked before each turn. - **With a framework alone**: A budget the team sets up and watches on its own. - **With an operating system**: A budget the system enforces by itself, before the money is spent. **In GrailAOS, today** - Usage measured at real cost and caps checked before each turn [Runs today] - Limit: The turn in progress always finishes; usage notifications are off by default and, like the billed-amount cap, need billing turned on. ### Rules enforced across the whole system The rules apply to the whole solution: they are verified before it runs, and the runtime enforces its guards while it runs. - **With a framework alone**: Validations the team adds wherever it needs them. - **With an operating system**: The same rules for every part, without the team having to repeat them in each one. **In GrailAOS, today** - The [compiler](https://grailaos.com/glossary#compiler) verifies the whole solution before anything runs [Runs today] - The notice of what the runtime does not execute [Partial] - Limit: The check covers a fixed set of cases. - Runtime guards: isolated [tenants](https://grailaos.com/glossary#tenant), integrations with no access to private networks and the contact's identity [Runs today] ### Health and observability of the whole You see the state of the whole system, not that of each agent separately, and someone finds out when something fails. - **With a framework alone**: Logs the team gathers and reviews on its own, unless it adds a platform. - **With an operating system**: The health and observability of the whole, with alerts to whoever operates it. **In GrailAOS, today** - Runtime health and the [readiness](https://grailaos.com/glossary#readiness) of each solution [Runs today] - One [trace](https://grailaos.com/glossary#trace) per turn and email alerts to whoever operates it [Runs today] - Traces in [OpenTelemetry](https://grailaos.com/glossary#opentelemetry) [Partial] - Limit: OpenTelemetry sends one span per agent turn: workflow runs are not sent. ### People in the loop A person steps in when needed and replies from within the same system, without leaving it. - **With a framework alone**: A pause for someone to approve; each team builds the inbox where a person replies. - **With an operating system**: People in the loop as part of the system: they step in when needed and the system keeps going. **In GrailAOS, today** - Handoff to a person, who replies from the console, and back to the agent [Runs today] - A person's approval as a step in a workflow [Runs today] ### Several agents as one unit Several agents and steps form a single workflow: in parallel, in a loop, with subworkflows and with a person who approves as one more step. - **With a framework alone**: A graph of agents the team programs; governing it together with the other pieces is up to the team. - **With an operating system**: Several agents and steps that form a single workflow, governed with the other pieces as one unit. **In GrailAOS, today** - Delegation between agents of the same solution [Runs today] - Conversational workflows [Runs today] - Workflows that run on their own, with list iteration, parallel branches, subworkflows and a person's approval as one more step [Runs today] ## For builders, buyers and investors The ladder helps you ask better questions. - **For builders**: Know which level you are at: for level 1 a framework is enough, for level 2 you need a platform that runs it, and level 3 calls for a different architecture. - **For buyers**: Ask which level the thing you are buying is at, and with what evidence: “multi-agent system” can mean almost anything. - **For investors**: What helps is knowing which level a product solves today, with what evidence, and what it is missing for the next one. - **For the industry**: It needs a shared vocabulary: with five levels, the conversation can be precise. ## Where GrailAOS stands We apply to GrailAOS the same rule we apply to the others in the industry lens, piece by piece and with the status of each one. - GrailAOS meets level 3 by Omniatix's criteria [Runs today] - Limit: Under other criteria, the result may change, and each piece has its limits: parallel branches run on a single server. GrailAOS is a [language](https://grailaos.com/glossary#language), a compiler and a runtime for programming businesses with agents, built with coding agents, and every statement about GrailAOS carries its status. We do not claim levels 4 and 5: they are the industry's horizon, and level 5 first requires advances in safety and in formal verification. - A language, a compiler and a runtime: the solution in versioned text documents, verified as a whole before it runs [Runs today] - Built with coding agents [Runs today] [The level 3 pieces, one by one](https://grailaos.com/papers/five-levels#why-os) **Measured in the repositories** - **v4.437** Specification version implemented by the runtime. Source: GrailAOS runtime repository, main branch. Measured on October 30, 2026. - **1,385** Specified features. Source: GrailAOS specification master index. Measured on October 30, 2026. - **564** Recorded architecture decisions. Source: GrailAOS specification master index. Measured on October 30, 2026. - **21,558** Tests declared in the runtime. Source: GrailAOS runtime repository, main branch. Measured on October 30, 2026. - **72.5%** Share of [commits](https://grailaos.com/glossary#commit) attributed to Claude. Source: GrailAOS specification and runtime repositories, main branch. Measured on October 30, 2026.