A look at the state of the AI agent industry
This autumn’s releases from OpenAI, Anthropic and Manus look unrelated. Read together, they show an industry that has agreed on how to keep an agent in check, and has built it one vendor at a time.
By the STACK team.
Since the end of September, the companies that build AI agents have put out a run of products that look, taken one at a time, like routine updates. OpenAI gave ChatGPT subscribers agents that keep working after the user has logged off, and on 13 November it gave those agents a way to pay by card. A day earlier, Anthropic said that companies can now run its agent service on their own servers. Manus, meanwhile, is selling an app in which every agent has its own phone number, email address and wallet.
Read together, the releases say something about where the industry is heading. The three companies have arrived at much the same way of keeping an agent in check, and each has built it for its own agents only. This article looks at what they released, why they built it the way they did, and what is still missing.
We make STACK, which is built along the same lines and is not tied to any of the three, so we have an interest in the argument. The last part of the article says where we come in.
Why an agent has to be kept in check
An agent is useful because it can act. It sends the email, books the room and pays the invoice. To do that it needs things a chat assistant never had, such as the login to a mailbox or the details of a card.
It also reads a great deal of text written by strangers, and some of that text is written to give it orders. An email can tell an agent to forward the whole mailbox to another address. An agent that obeys is, for that moment, working for whoever wrote the email. The technique is called prompt injection.
There are two ways to deal with it. One is to make the model careful and trust it to refuse. The other is to assume that the model will sometimes be fooled, and to arrange things so that a fooled agent cannot do much. The products released this autumn all take the second way.
Three companies, one approach
Anthropic is the most explicit about it. Claude Managed Agents is a service for developers who build agents on Claude, and the company’s own examples are agents that fix broken builds, test releases and work through a backlog of bugs with nobody watching. In its announcement Anthropic says that because no person can check every step of such an agent, “the limits have to be built into the system itself”.
In practice that means four things. The agent’s code runs in a sandbox, separate from the loop that calls the model. Logins are held in a vault outside the sandbox, and a proxy adds them to a request on its way out, so the agent never holds one. Each tool can be set to run freely, to wait for a person’s approval, or to be judged call by call by Anthropic’s server. And every step of a session is recorded. With this week’s release a company can run all of it inside its own network.
OpenAI’s dots are aimed at a different buyer, the ChatGPT subscriber, and they reach the same place. A dot has its own cloud computer and browser and connects to more than 4,000 apps. It can sign in to websites with saved passwords that are never shown to the model. The owner writes rules that allow an action, require approval for it or block it, and some actions, such as changing a password, are never left to the dot. A monitoring system can pause or stop a dot’s work.
The wallet OpenAI added on 13 November follows the same thinking. The user adds a card once. For each purchase the card network issues a substitute number, and the user approves the purchase with a fingerprint or a face check. Neither the dot nor the shop sees the real card. OpenAI built it with American Express, Mastercard and Visa, and it is available first to Pro subscribers in the United States.
Manus’s Cue goes furthest in what it gives the agent. Its page describes an agent with a phone number that takes the calls you forward to it, an email address of its own, a card to pay with and a computer to work on, with tools and permissions set for each agent. The page does not say how those permissions are enforced, or what record is kept.
Set side by side, the parts are largely the same.
| Claude Managed Agents | OpenAI dots | Manus Cue | |
|---|---|---|---|
| Logins and cards kept from the model | A vault, with a proxy that adds the login to each request | Saved passwords the model never sees. A substitute card number for each purchase | Not stated on its page |
| Rules checked outside the model | Each tool runs freely, waits for approval, or is judged call by call | Rules that allow, require approval or block. Every purchase approved by the user | Tools and permissions set for each agent |
| A record | Every step of a session | An activity view, and a record of each approved purchase | Not stated on its page |
| An identity for the agent | A signed token, on self-hosted servers | Announced for specialist dots, as a preview | A phone number and an email address |
| What the agent is given | Nothing beyond its tools | A card | A phone number, an email address and a card |
| Where the agent runs | A sandbox | A cloud computer | A computer workspace |
| Works with | Agents built on Claude | OpenAI’s dots | Agents inside Cue |
A settled design
The table shows a good deal of agreement for three companies that compete with each other. Anthropic and OpenAI do not ask the model to guard a password or to decide what it is allowed to do. Those jobs are done by software around the model, which holds the logins, applies the rules and keeps the record. Cue’s page says less about how it works, but what it describes points the same way.
Looked at closely, there are two layers here. The lower one keeps the agent in check: it holds the logins, applies the rules and keeps the record. The upper one is what the agent is given to work with, such as a card, a phone number or an email address. Anthropic has built the lower layer and nothing on top of it. OpenAI has built both, and added the card on 13 November. Manus shows the upper layer, and its page says little about what lies underneath.
Built once for each vendor
Each company has built this inside its own product, which is natural, since each started from its own agent. Anthropic’s vault and rules apply to agents built on Managed Agents, OpenAI’s rules and wallet apply to dots, and Cue’s phone number and card belong to agents inside Cue.
The things being held, though, are not the vendor’s. The mailbox, the card, the phone number and the record of what was done all belong to the person or company that owns the agent, and that owner is unlikely to stay with one agent. The three products are sold for different jobs. Anthropic’s examples are agents that work on code, while OpenAI and Manus show agents that book travel, answer calls and send invoices, so a company that wants both kinds is already looking at two products. Models change as well: OpenAI added a new model to dots five weeks after launching them. And some agents run on no vendor’s service at all. LM Studio now ships Bionic, an agent for open models that can run on the user’s own computer.
The vendors’ own choices suggest that the layer can stand apart from the model. Anthropic connects tools through MCP, exports the record over OpenTelemetry and signs people in through OpenID Connect, all of them open standards. OpenAI built its wallet with the card networks so that it works at an ordinary checkout, and its guidance to shops is about accepting purchases made by agents in general.
The case for putting it with the owner
If the design is settled, and what it protects belongs to the owner, the remaining question is where the layer should sit. Inside each agent’s product, it has to be set up again for every agent, and the logins, the limits and the record are split between them. With the owner, it is set up once and every agent uses it.
That is what STACK is. It has both layers, and both belong to the owner. The lower one holds the logins, checks each request and keeps the record. The upper one is an account with an email address, a phone number, a card, a wallet and a postal address, and everything in it passes through the lower layer. We do not make a model and we do not run the agent. An agent connects from wherever it runs, on whichever model, and several agents share the account.
When the owner changes model, or adds an agent from another vendor, the logins, the limits and the record stay where they are. This autumn’s releases show that the largest companies in the field now build agents this way. What they have not built is a version that belongs to the owner, and that is the one we make.
Sources
- Anthropic: self-hosted Claude Managed Agents, 12 November 2026
- Anthropic: Claude Managed Agents
- Anthropic docs: vaults
- Anthropic docs: permission policies
- Anthropic docs: differences from the hosted service
- OpenAI: Introducing dots, 29 September 2026
- OpenAI: Introducing ChatGPT Wallet, 13 November 2026
- Manus: Cue
- LM Studio: Bionic
- STACK: every service and its price
- STACK: how controls work
- STACK: integrations
STACK is not affiliated with Anthropic, OpenAI or Manus. “Not stated on its page” means the company’s page does not say. To correct anything here, write to hello@getstack.run.