Anthropic's One Claude: Make Automatic Capability Routing Visible

What Anthropic changed on 16 September

Anthropic said Claude chat and Cowork are merging into one interface. Claude Docs and Claude Slides launched in beta, while Claude Design can now work inside ordinary conversations. The company says Claude can determine what a task needs and apply Cowork or Design capabilities with the context, skills and connectors already available to the user.

The first rollout is to Pro and Max plans across web, desktop and mobile over the coming weeks. Team and Free plans are due later. Anthropic says Enterprise administrators choose when to enable Docs, Slides and Design, and will receive at least 30 days' notice before the broader interface change reaches their organisations.

Outputs can stay in Claude as shareable documents and presentations. Anthropic also says users can export documents to Google Docs or Microsoft Word and presentations to PowerPoint or PDF. Reuters and The Verge independently reported the unified interface, automatic capability selection and rollout sequence.

A simpler interface creates a hidden routing decision

The productivity gain is real: a person can ask for analysis, a report and a leadership deck without deciding which Claude surface should handle each step. The control question moves rather than disappears. Instead of a user explicitly selecting chat, an agentic workspace or a creation tool, the system interprets the task and chooses the capability path.

That path can affect which connectors are called, how long work continues, whether files are created, how collaborators gain access and where the final artefact is exported. Anthropic says Claude asks before taking an action by default and lets users change how often it checks in. An organisation still needs its own rules for when that discretion is acceptable and when a human must approve the route or output.

Require a capability receipt for consequential work

Treat automatic routing as an operational decision that should leave evidence. For each consequential task, capture a compact receipt with five fields:

• Intent: the user's request, business purpose and authorised output class.
• Route: the capabilities Claude selected, including Cowork, Docs, Slides, Design, skills and connectors.
• Data: the sources read, files created, destinations used and access boundaries applied.
• Control: the action mode, approval points, collaborator permissions and person accountable for release.
• Result: the final artefact, export format, material edits, exceptions and verification outcome.

Pilot the workflow, not just the interface

Start with one repeatable, reversible workflow such as a weekly operating report and slide summary. Define the approved connectors, destination folders, export formats and sign-off owner before the pilot. Run representative prompts, including ambiguous requests, unavailable sources and conflicting instructions, then compare Claude's selected route with the route an experienced operator expected.

Measure more than completion time. Track route accuracy, unsupported claims, unnecessary connector use, manual corrections, approval latency and export defects. Keep consequential external actions behind an approval gate, and use observable action logs rather than private reasoning as the audit record.

ELYMENT AI helps teams turn unified assistants into governed operating workflows with explicit authority, evidence and human review. The practical goal is to preserve the speed of one interface without losing visibility into the different capabilities acting underneath it.

Sources

• Anthropic, Claude Cowork and chat are now one Claude (16 September 2026) - Primary announcement covering the unified interface, automatic task handling, Docs, Slides, Design, approvals and rollout.
• Reuters, Anthropic to fold Claude AI features into one interface (16 September 2026) - Independent reporting on the product changes, export options and plan sequence.
• The Verge, Claude comes for Gemini with its own take on Docs and Slides (16 September 2026) - Independent product reporting on collaborative editing, shareable links, capability access and rollout.

Continue learning

• Design human approval into agent workflows
• Build visible evidence for agent actions
• Audit AI actions rather than private reasoning

Frequently asked questions