Short answer
An agent OS (agent operating system) is the software layer that runs AI agents for a business. It holds each agent's instructions, tools, memory and permissions, and records what every agent did, the way an operating system manages programs.- Also called
- AgentOS, AI agent OS, agent operating system
- Core parts
- Instructions, tools, memory, permissions, orchestration, logs
- Not the same as
- An agent framework, which is code a developer builds agents with
- The test
- Can you see and limit what each agent is allowed to do?
The comparison to a computer's operating system is deliberate. A computer's OS decides which program gets memory, which files it can open and which devices it can reach. An agent OS does the same for AI agents: which agent handles a request, what it is allowed to read, which actions it may take, and what happens when it should stop and hand the work to a person.
What an agent OS does
Most products sold as an agent OS cover the same six jobs.
- Instructions. The written rules each agent follows, versioned so a change can be reviewed and undone.
- Tools and data. Connections to the systems an agent reads and acts on, such as an order system, a CRM or a helpdesk.
- Memory. What the agent keeps between turns and between conversations, and what it is never allowed to keep.
- Permissions. Which agent may take which action, and which actions need a person to approve them first.
- Orchestration. Routing a request to the right agent, or passing it between agents or to a person.
- Logs. A record of every step: what the agent read, what it decided and what it changed.
The last two are where most of the difference lies. Running an agent is easy. Knowing what it did, and limiting what it may do, is the hard part.
In Aide, instructions are versioned this way: an edit to a live procedure stays a draft while the live version keeps running, Set live takes a note, and version history shows who changed what and can restore an older version. The workspace audit log records every change to agents, procedures, topics and knowledge, with who made it and a before and after.
Agent OS vs agent framework vs chatbot platform
| Agent OS | Agent framework | Chatbot platform | |
|---|---|---|---|
| What it is | A runtime and control layer for many agents | A code library for building an agent | A builder for scripted conversation flows |
| Who uses it | Operations and IT teams | Developers | Support and marketing teams |
| How agents act | Through connected tools, inside permissions | However the developer codes it | They mostly answer, rarely act |
| Visibility | Logs and approvals built in | Whatever the developer adds | Conversation transcripts |
A framework is something you build with. An agent OS is something you run agents on. A chatbot platform builds flows that answer, and usually stops short of acting.
How to evaluate an agent OS
The label covers very different products, so judge one by what it shows you, not by the name.
- Permissions per action. Can you allow an agent to look up an order but not refund it?
- Testing before live. Can you try a change on real past requests before customers see it?
- Logs a person can read. Does every action show what the agent read and why it acted?
- A clear way out. When the agent is unsure or the request is out of scope, does it hand off with context, or guess?
- Scope you can grow. Can you start with one kind of request and add more as each one proves itself?
The pitch this page rejects is the one that measures an agent OS by how many agents it runs. More agents with less visibility is more risk, not more capability.
With Aide, the first test looks like this: Aide can look up and cancel a verified shopper's Shopify order, while every refund waits for a teammate by default. A large refund started by a teammate with the Agent role waits in turn for an owner or admin to approve.
Where Aide fits
Aide, the AI platform for customer service, is built around the customer request rather than the agent. Every message is sorted into the team's own topics, and each topic is set to answer and act on its own, write drafts for a person, or stay with the team. Situations that need steps get a procedure the team writes and checks against its last 50 conversations before setting it live, and every reply, source and order action is recorded on the conversation. See how the agent itself works on the AI agents page.