Editorial note: Independent desk analysis based on the primary or original reporting sources listed below. No sponsor reviewed or paid for this article.

Direct answer: Reuters reported on September 16 that Anthropic is unifying Claude chat and Cowork into one interface and adding Claude Docs, Slides and Design capabilities. The UX direction is clear: instead of making users choose the correct AI mode before starting, the system increasingly selects capabilities from the task itself.

Mode selection is product taxonomy leaking into the UI

AI products accumulated modes for understandable reasons. Chat, research, coding, image creation and computer-use workflows often have different models, permissions, latency and toolchains. A mode picker exposes those differences directly to the user.

The problem is that users rarely arrive with the platform’s taxonomy in mind. They arrive with an outcome: compare these documents, make the presentation, redesign this asset, analyze these files. Every time the interface asks which internal mode should handle the task, it transfers routing work from the system to the user.

Anthropic is collapsing that routing step

Reuters reported that Anthropic is bringing Claude’s chat and Cowork features into one interface and introducing Docs, Slides and Design tools, with Claude determining which capabilities are needed. The company’s Cowork direction has already emphasized multi-step workflows rather than a single conversational response.

The significance is not that every task becomes chat. It is almost the opposite. Conversation can remain the intent surface while the system invokes structured creation, file operations or visual tooling behind it.

Automatic routing needs visible execution state

Removing the mode picker creates a different UX obligation: the product must make the chosen execution path legible. If the system silently switches from answering a question to editing files, operating software or preparing an export, the user needs to understand that state change.

The interface should reveal the capability being used when it affects permissions, cost, latency or side effects. Good routing is invisible when the distinction does not matter and explicit when it does. That is a more useful rule than either “show every tool” or “hide all implementation detail.”

The prompt becomes an intent contract, not a tool command

A unified interface works best when the user can state success criteria rather than implementation instructions. “Turn these notes into a six-slide investor update and export a PowerPoint” is a stronger interaction than asking the user to first choose a slide mode, then a document source mode, then an export path.

For product teams, this shifts design work toward task decomposition and state transitions. The system needs to preserve context as it moves across capabilities, ask for approval at consequential boundaries and return an artifact that remains editable after generation.

The winning AI interface may have fewer modes and more states

Mode-heavy products often look powerful because they expose many capabilities. State-rich products feel powerful because the user can understand what is happening: gathering inputs, planning, generating, editing, waiting for approval, exporting, recovering.

As AI systems become better at routing themselves, the visible product surface can become simpler without becoming less controllable. The competitive design problem is no longer how many AI buttons fit in the toolbar. It is how little orchestration the user has to manage while still retaining control over consequential work.

When a mode should remain visible

Unified does not have to mean indistinguishable. A visible mode is still useful when choosing it changes the user’s commitment: a long research run may consume more time or budget; computer use may require broader permissions; design generation may create new files; a coding session may execute untrusted output. In those cases, the product can infer the route and then show a concise transition before work begins. The system carries the routing burden while the user retains authority over the consequence.

The same principle applies to professional identity. A user may want an exploratory answer in one moment and a traceable, reproducible artifact in another. The interface should infer as much as it safely can from the requested outcome, then expose the few decisions that materially change quality, cost, permissions or deliverables.

How to evaluate automatic routing

Routing accuracy alone is not enough. Teams should measure correction rate, unnecessary clarification, permission surprises, task switching, artifact quality and recovery after a wrong route. A successful router does not merely select the intended tool; it notices when evidence changes and can move to another capability without discarding context. The UX benchmark is whether users spend less time managing the AI’s internal structure while maintaining the same or greater control over the result.

Practical takeaways

  • Route by user outcome whenever internal mode differences do not materially affect the user.
  • Reveal capability changes when they affect permissions, cost, latency or side effects.
  • Preserve context across chat, document, design and execution states.
  • Ask for approval at consequential transitions rather than at every internal tool call.
  • Design for clear states and recoverability instead of exposing a growing tool taxonomy.

Related reading

Sources