Claude Agents Map: Six Products, Two Axes, and What Each Is Published to Do

Search interest in the exact query claude agents sits at an index of 10 in the seven days to 23 October 2026, up 30% on the period before (Google Trends, worldwide, past seven days, read on 23 October 2026). Small numbers, clear direction, and a predictable next step for anyone who follows the query: a navigation menu with six product names on it and no obvious order. This piece is the ordering.

Six products carry agent behaviour at Anthropic: Claude Code, Claude Cowork, Claude Duet, Claude Tag, Claude Science and Claude Security. Each has its own page, its own availability line and its own job. Read as a set, two questions sort them faster than any feature list: does it wait for you or come at you, and where does it live.

Six names on one navigation bar

Anthropic's product navigation lists Claude, then the six above. The plain Claude entry is the chat surface; the other six are the ones whose descriptions are written in terms of tasks. All six are Anthropic products, and none of them is a model you call.

The names do not share a scheme, which is why a map is needed at all. Code and Cowork name a mode of work. Duet names a kind of relationship. Tag borrows a convention from chat. Science and Security name a field. A reader who knows what Claude Code does has no way to infer what Claude Tag does from the naming, and the product pages do not cross-reference each other. The only way to hold the set in your head is to sort it by something the names are not telling you.

Axis one: where it lives

The first axis is surface. Claude Code is the widest: terminal, desktop app, web interface, mobile app, with GitHub and popular editor integrations alongside. Claude Cowork runs on web, desktop and mobile, and works inside folders you choose. Claude Duet is a standalone app. Claude Tag lives in Slack. Claude Science is a desktop research app for macOS, Windows and Linux. Claude Security, in its published description, is a public beta for Claude Enterprise that scans with Claude Mythos; the surface an administrator works in is not published, and the published description does not name one.

The Claude Code page is a fair illustration of how much surface one product can span: a download button, an availability line naming three operating systems, and a row of other ways to reach the same product, including a terminal, a browser, two phone platforms, GitHub, two editors and Slack. Six products with that kind of spread cannot be compared on surface alone, which is why the second axis carries more weight.

Axis two: pull and push

The second axis is who starts the work. Five of the six are pull. Claude Code takes a bug fix, a test or a multi-day migration that you steer and review. Claude Cowork takes a goal and works across your files and tools while you follow along and redirect. Claude Tag reacts inside a thread somebody else began. Claude Science and Claude Security are pointed at a body of work, a research question or a codebase to scan, and both are described by their output rather than by when they interrupt you.

Pull does not mean passive. Cowork's page says you can schedule a task for any cadence and it runs unattended, and that closing your laptop does not stop it. The distinction is not about whether the work is continuous, it is about who decides that the work starts. A schedule you set is still you pulling.

Claude Duet is the push side. Anthropic describes it as a proactive agent: a standalone app on a waitlist for Claude Pro and Max subscribers in the US, built on Claude Opus 5.5. Proactive is the load-bearing word. It is the only product in the set whose published description is not organised around a task you hand it, and that single difference is what makes the other five look like tools and this one look like a service.

What that means in practice is also the least documented thing on the map. Whether Duet reaches users outside the US is not published; the waitlist is the published route in, and no date for another country appears beside it. A reader planning around Duet in October 2026 is planning around a research preview, not around a rollout.

Code and Cowork: the pair the map exists to separate

The two most likely to be confused are the two that share the most. Anthropic's own Cowork page answers the question directly. Claude Code is built for software engineering, for writing, debugging and shipping code, and is available as a terminal tool, a desktop app, a web interface and a mobile app, with GitHub and popular IDE integrations. Cowork is built for non-coding knowledge work: research, analysis, document creation and other multi-step tasks. The page's own framing is that Cowork uses the same agentic approach as Claude Code.

That sentence is the most useful one on either page, because it separates the products by domain rather than by capability. The agent loop is the same; the work is not. Code's examples are bug fixes, tests and migrations. Cowork's are folders of exports, spreadsheets, contract review and scheduled reports. A team that buys one expecting the other has bought a capable product for the wrong job, which is an expensive way to learn the distinction.

The control story differs too, and it is the part a buyer should read closely. Cowork works in the folders and tools you choose, opens a browser in its own side panel rather than yours, can be configured to show its plan and wait for approval before significant actions, and needs your approval to delete anything. Code's page makes a related but differently shaped claim: used in a terminal, it runs locally and talks directly to model APIs without a backend server or remote code index, and it asks for permission before changing your files or running commands.

Two products, two answers to what it can touch and who approves it. Both answers are worth reading before a rollout rather than after, because they describe different permission models for what is, underneath, the same agent loop.

The three scoped ones: Tag, Science, Security

The remaining three are narrower, and the narrowness is the clearest thing about them. Claude Tag is Claude in Slack, available for Claude Enterprise and Team customers in Slack. Claude Science is a beta desktop research app that produces reproducible artifacts and runs on macOS, Windows and Linux. Claude Security is a public beta for Claude Enterprise that scans with Claude Mythos.

All three are gated by something other than a purchase decision. Tag is gated by plan and by the fact that your organisation works in Slack. Security is gated by an Enterprise plan. Science is gated by being a beta, which means its published description is a statement of where the product is going as much as what it does today. If you are not in one of those situations, these are not products you can evaluate yet. They are products you can read about, and the reading is short.

What the map does not publish

A map is also a record of blank space, and this one has several. The most conspicuous is that Anthropic publishes no single page placing the six side by side. Every comparison in this piece had to be assembled from six product pages, and that assembly is the map.

The blanks are specific. A date for Claude Duet outside the US is not published; the published route is a waitlist for Pro and Max subscribers in the US. A general-availability date for Claude Science is not published, and a general-availability date for Claude Security is not published either; both are described as betas, which is a state rather than a schedule. Where Claude Security's scanning happens, beyond a public beta for Claude Enterprise, is not published. And no page in the set describes how much the six share underneath: whether two of them can pick up the same task is a question the pages do not answer.

None of that is a criticism. Product pages are written for buyers of one product, not for readers comparing six, and the absence of a comparison page is normal. It does mean that anyone who needs the comparison has to do what this article did, which is read all six and notice which sentences are doing the sorting.

What this map is not

We host models, not any of these apps. OrcaRouter does not sell Claude Code, Claude Cowork, Claude Duet, Claude Tag, Claude Science or Claude Security, and no endpoint of ours produces one of them. The distinction matters more here than usual, because agent is a word used for both halves of the stack: the model that reasons, and the product that decides what to point it at and when.

The half we sell is the model behind one OpenAI-compatible interface and one key, across 200-plus models at provider list price with no markup, with automatic failover when a provider has a bad minute. If you are building something on the pattern this map describes, that is the part you can buy. The surface, the permissions and the judgement about when to act are the part you build, and each of the six products above is one answer to how that part should look.

The map in one reading

Read the menu again with the two axes in mind and it stops being a list of names. Code is pull, and it meets developers everywhere they already work. Cowork is pull, on your files, on every screen you own. Duet is push, a standalone app in the US on a waitlist. Tag is pull, inside Slack, for Enterprise and Team. Science is pull, a beta research app on a lab desktop. Security is pull, inside an Enterprise plan, scanning with Claude Mythos.

Six products, one agent loop, and a set of decisions about surface and initiation that turn out to be the actual difference between them. That is what the word agents covers when a search brings you to Anthropic's menu, and it is more useful than the menu itself, because the menu only tells you the names.